Computer Vision – ECCV 2020

LiDAR point-cloud segmentation is an important problem for many applications. For large-scale point cloud segmentation, the de facto method is to project a 3D point cloud to get a 2D LiDAR image and use convolutions to process it. Despite the similarity between regular RGB and LiDAR images, we are the first to discover that the feature distribution of LiDAR images changes drastically at different image locations. Using standard convolutions to process such LiDAR images is problematic, as convolution filters pick up local features that are only active in specific regions in the image. As a result, the capacity of the network is under-utilized and the segmentation performance decreases. To fix this, we propose Spatially-Adaptive Convolution (SAC) to adopt different filters for different locations according to the input image. SAC can be computed efficiently since it can be implemented as a series of element-wise multiplications, im2col, and standard convolution. It is a general framework such that several previous methods can be seen as special cases of SAC. Using SAC, we build SqueezeSegV3 for LiDAR point-cloud segmentation and outperform all previous published methods by at least 2.0% mIoU on the SemanticKITTI benchmark. Code and pretrained model are available at https://github.com/chenfengxu714/SqueezeSegV3 .

The progressive clustering method and its variants, which iteratively generate pseudo labels for unlabeled data and per form feature learning, have shown great process in unsupervised person re-identification (re-id). However, they have an intrinsic problem of modeling the in-camera variability of images successfully, that is, pedestrian features extracted from the same camera tend to be clustered into the same class. This often results in a non-convergent model in the real world application of clustering based re-id models, leading to degenerated performance. In the present study, we propose an attention-driven two-stage clustering (ADTC) method to solve this problem. Specifically, our method consists of two strategies. Firstly, we use an unsupervised attention kernel to shift the learned features from the image background to the pedestrian foreground, which results in more informative clusters. Secondly, to aid the learning of the attention driven clustering model, we separate the clustering process into two stages. We first use kmeans to generate the centroids of clusters (stage 1) and then apply the k-reciprocal Jaccard distance (KRJD) metric to re-assign data points to each cluster (stage 2). By iteratively learning with the two strategies, the attentive regions are gradually shifted from the background to the foreground and the features become more discriminative. Using two benchmark datasets Market1501 and DukeMTMC, we demonstrate that our model outperforms other state-of-the-art unsupervised approaches for person re-id.

Facial expression manipulation aims at editing facial expression with a given condition. Previous methods edit an input image under the guidance of a discrete emotion label or absolute condition (e.g., facial action units) to possess the desired expression. However, these methods either suffer from changing condition-irrelevant regions or are inefficient for fine-grained editing. In this study, we take these two objectives into consideration and propose a novel method. First, we replace continuous absolute condition with relative condition, specifically, relative action units. With relative action units, the generator learns to only transform regions of interest which are specified by non-zero-valued relative AUs. Second, our generator is built on U-Net but strengthened by multi-scale feature fusion (MSF) mechanism for high-quality expression editing purposes. Extensive experiments on both quantitative and qualitative evaluation demonstrate the improvements of our proposed approach compared to the state-of-the-art expression editing methods. Code is available at https://github.com/junleen/Expression-manipulator .

In this paper, we propose a novel end-to-end unsupervised deep domain adaptation model for adaptive object detection by exploiting multi-label object recognition as a dual auxiliary task. The model exploits multi-label prediction to reveal the object category information in each image and then uses the prediction results to perform conditional adversarial global feature alignment, such that the multimodal structure of image features can be tackled to bridge the domain divergence at the global feature level while preserving the discriminability of the features. Moreover, we introduce a prediction consistency regularization mechanism to assist object detection, which uses the multi-label prediction results as an auxiliary regularization information to ensure consistent object category discoveries between the object recognition task and the object detection task. Experiments are conducted on a few benchmark datasets and the results show the proposed model outperforms the state-of-the-art comparison methods .

Tables are information-rich structured objects in document images. While significant work has been done in localizing tables as graphic objects in document images, only limited attempts exist on table structure recognition. Most existing literature on structure recognition depends on extraction of meta-features from the pdf document or on the optical character recognition (ocr) models to extract low-level layout features from the image. However, these methods fail to generalize well because of the absence of meta-features or errors made by the ocr when there is a significant variance in table layouts and text organization. In our work, we focus on tables that have complex structures, dense content, and varying layouts with no dependency on meta-features and/or ocr.We present an approach for table structure recognition that combines cell detection and interaction modules to localize the cells and predict their row and column associations with other detected cells. We incorporate structural constraints as additional differential components to the loss function for cell detection. We empirically validate our method on the publicly available real-world datasets - icdar-2013, icdar-2019 (ctdar) archival, unlv, scitsr, scitsr-comp, tablebank, and pubtabnet. Our attempt opens up a new direction for table structure recognition by combining top-down (table cells detection) and bottom-up (structure recognition) cues in visually understanding the tables.

Novel view synthesis often needs the paired data from both the source and target views. This paper proposes a view translation model under cVAE-GAN framework without requiring the paired data. We design a conditional deformable module (CDM) which uses the view condition vectors as the filters to convolve the feature maps of the main branch in VAE. It generates several pairs of displacement maps to deform the features, like the 2D optical flows. The results are fed into the deformed feature based normalization module (DFNM), which scales and offsets the main branch feature, given its deformed one as the input from the side branch. Taking the advantage of the CDM and DFNM, the encoder outputs a view-irrelevant posterior, while the decoder takes the code drawn from it to synthesize the reconstructed and the view-translated images. To further ensure the disentanglement between the views and other factors, we add adversarial training on the code. The results and ablation studies on MultiPIE and 3D chair datasets validate the effectiveness of the framework in cVAE and the designed module.

We develop a language-guided navigation task set in a continuous 3D environment where agents must execute low-level actions to follow natural language navigation directions. By being situated in continuous environments, this setting lifts a number of assumptions implicit in prior work that represents environments as a sparse graph of panoramas with edges corresponding to navigability. Specifically, our setting drops the presumptions of known environment topologies, short-range oracle navigation, and perfect agent localization. To contextualize this new task, we develop models that mirror many of the advances made in prior settings as well as single-modality baselines. While some transfer, we find significantly lower absolute performance in the continuous setting – suggesting that performance in prior ‘navigation-graph’ settings may be inflated by the strong implicit assumptions. Code at jacobkrantz.github.io/vlnce .

Temporal action proposal generation plays an important role in video action understanding, which requires localizing high-quality action content precisely. However, generating temporal proposals with both precise boundaries and high-quality action content is extremely challenging. To address this issue, we propose a novel Boundary Content Graph Neural Network (BC-GNN) to model the insightful relations between the boundary and action content of temporal proposals by the graph neural networks. In BC-GNN, the boundaries and content of temporal proposals are taken as the nodes and edges of the graph neural network, respectively, where they are spontaneously linked. Then a novel graph computation operation is proposed to update features of edges and nodes. After that, one updated edge and two nodes it connects are used to predict boundary probabilities and content confidence score, which will be combined to generate a final high-quality proposal. Experiments are conducted on two mainstream datasets: ActivityNet-1.3 and THUMOS14. Without the bells and whistles, BC-GNN outperforms previous state-of-the-art methods in both temporal action proposal and temporal action detection tasks.

Object pose increases intraclass object variance which makes object recognition from 2D images harder. To render a classifier robust to pose variations, most deep neural networks try to eliminate the influence of pose by using large datasets with many poses for each class. Here, we propose a different approach: a class-agnostic object pose transformation network (OPT-Net) can transform an image along 3D yaw and pitch axes to synthesize additional poses continuously. Synthesized images lead to better training of an object classifier. We design a novel eliminate-add structure to explicitly disentangle pose from object identity: first ‘eliminate’ pose information of the input image and then ‘add’ target pose information (regularized as continuous variables) to synthesize any target pose. We trained OPT-Net on images of toy vehicles shot on a turntable from the iLab-20M dataset. After training on unbalanced discrete poses (5 classes with 6 poses per object instance, plus 5 classes with only 2 poses), we show that OPT-Net can synthesize balanced continuous new poses along yaw and pitch axes with high quality. Training a ResNet-18 classifier with original plus synthesized poses improves mAP accuracy by 9% over training on original poses only. Further, the pre-trained OPT-Net can generalize to new object classes, which we demonstrate on both iLab-20M and RGB-D. We also show that the learned features can generalize to ImageNet. (The code is released at this github url ).

Video Moment Retrieval (VMR) is a task to localize the temporal moment in untrimmed video specified by natural language query. For VMR, several methods that require full supervision for training have been proposed. Unfortunately, acquiring a large number of training videos with labeled temporal boundaries for each query is a labor-intensive process. This paper explores a method for performing VMR in a weakly-supervised manner (wVMR): training is performed without temporal moment labels but only with the text query that describes a segment of the video. Existing methods on wVMR generate multi-scale proposals and apply query-guided attention mechanism to highlight the most relevant proposal. To leverage the weak supervision, contrastive learning is used which predicts higher scores for the correct video-query pairs than for the incorrect pairs. It has been observed that a large number of candidate proposals, coarse query representation, and one-way attention mechanism lead to blurry attention map which limits the localization performance. To address this issue, Video-Language Alignment Network (VLANet) is proposed that learns a sharper attention by pruning out spurious candidate proposals and applying a multi-directional attention mechanism with fine-grained query representation. The Surrogate Proposal Selection module selects a proposal based on the proximity to the query in the joint embedding space, and thus substantially reduces candidate proposals which leads to lower computation load and sharper attention. Next, the Cascaded Cross-modal Attention module considers dense feature interactions and multi-directional attention flows to learn the multi-modal alignment. VLANet is trained end-to-end using contrastive loss which enforces semantically similar videos and queries to cluster. The experiments show that the method achieves state-of-the-art performance on Charades-STA and DiDeMo datasets.

Query expansion is a technique widely used in image search consisting in combining highly ranked images from an original query into an expanded query that is then reissued, generally leading to increased recall and precision. An important aspect of query expansion is choosing an appropriate way to combine the images into a new query. Interestingly, despite the undeniable empirical success of query expansion, ad-hoc methods with different caveats have dominated the landscape, and not a lot of research has been done on learning how to do query expansion. In this paper we propose a more principled framework to query expansion, where one trains, in a discriminative manner, a model that learns how images should be aggregated to form the expanded query. Within this framework, we propose a model that leverages a self-attention mechanism to effectively learn how to transfer information between the different images before aggregating them. Our approach obtains higher accuracy than existing approaches on standard benchmarks. More importantly, our approach is the only one that consistently shows high accuracy under different regimes, overcoming caveats of existing methods.

This paper proposes a knowledge distillation method for foreground object search (FoS). Given a background and a rectangle specifying the foreground location and scale, FoS retrieves compatible foregrounds in a certain category for later image composition. Foregrounds within the same category can be grouped into a small number of patterns. Instances within each pattern are compatible with any query input interchangeably. These instances are referred to as interchangeable foregrounds. We first present a pipeline to build pattern-level FoS dataset containing labels of interchangeable foregrounds. We then establish a benchmark dataset for further training and testing following the pipeline. As for the proposed method, we first train a foreground encoder to learn representations of interchangeable foregrounds. We then train a query encoder to learn query-foreground compatibility following a knowledge distillation framework. It aims to transfer knowledge from interchangeable foregrounds to supervise representation learning of compatibility. The query feature representation is projected to the same latent space as interchangeable foregrounds, enabling very efficient and interpretable instance-level search. Furthermore, pattern-level search is feasible to retrieve more controllable, reasonable and diverse foregrounds. The proposed method outperforms the previous state-of-the-art by $$10.42\%$$ in absolute difference and $$24.06\%$$ in relative improvement evaluated by mean average precision (mAP). Extensive experimental results also demonstrate its efficacy from various aspects. The benchmark dataset and code will be release shortly.

Most previous knowledge distillation frameworks train the student to mimic the teacher’s output of each sample or transfer cross-sample relations from the teacher to the student. Nevertheless, they neglect the structured relations at a category level. In this paper, a novel Category Structure is proposed to transfer category-level structured relations for knowledge distillation. It models two structured relations, including intra-category structure and inter-category structure, which are intrinsic natures in relations between samples. Intra-category structure penalizes the structured relations in samples from the same category and inter-category structure focuses on cross-category relations at a category level. Transferring category structure from the teacher to the student supplements category-level structured relations for training a better student. Extensive experiments show that our method groups samples from the same category tighter in the embedding space and the superiority of our method in comparison with closely related works are validated in different datasets and models.

Generating photorealistic images of human faces at scale remains a prohibitively difficult task using computer graphics approaches. This is because these require the simulation of light to be photorealistic, which in turn requires physically accurate modelling of geometry, materials, and light sources, for both the head and the surrounding scene. Non-photorealistic renders however are increasingly easy to produce. In contrast to computer graphics approaches, generative models learned from more readily available 2D image data have been shown to produce samples of human faces that are hard to distinguish from real data. The process of learning usually corresponds to a loss of control over the shape and appearance of the generated images. For instance, even simple disentangling tasks such as modifying the hair independently of the face, which is trivial to accomplish in a computer graphics approach, remains an open research question. In this work, we propose an algorithm that matches a non-photorealistic, synthetically generated image to a latent vector of a pretrained StyleGAN2 model which, in turn, maps the vector to a photorealistic image of a person of the same pose, expression, hair, and lighting. In contrast to most previous work, we require no synthetic training data. To the best of our knowledge, this is the first algorithm of its kind to work at a resolution of 1K and represents a significant leap forward in visual realism.

Few-shot learning, namely recognizing novel categories with a very small amount of training examples, is a challenging area of machine learning research. Traditional deep learning methods require massive training data to tune the huge number of parameters, which is often impractical and prone to over-fitting. In this work, we further research on the well-known few-shot learning method known as prototypical networks for better performance. Our contributions include (1) a new embedding structure to encode relative spatial relationships between features by applying a capsule network; (2) a new triplet loss designated to enhance the semantic feature embedding where similar samples are close to each other while dissimilar samples are farther apart; and (3) an effective non-parametric classifier termed attentive prototypes in place of the simple prototypes in current few-shot learning. The proposed attentive prototype aggregates all of the instances in a support class which are weighted by their importance, defined by the reconstruction error for a given query. The reconstruction error allows the classification posterior probability to be estimated, which corresponds to the classification confidence score. Extensive experiments on three benchmark datasets demonstrate that our approach is effective for the few-shot classification task.

Recent approaches for weakly supervised instance segmentations depend on two components: (i) a pseudo label generation model which provides instances that are consistent with a given annotation; and (ii) an instance segmentation model, which is trained in a supervised manner using the pseudo labels as ground-truth. Unlike previous approaches, we explicitly model the uncertainty in the pseudo label generation process using a conditional distribution. The samples drawn from our conditional distribution provide accurate pseudo labels due to the use of semantic class aware unary terms, boundary aware pairwise smoothness terms, and annotation aware higher order terms. Furthermore, we represent the instance segmentation model as an annotation agnostic prediction distribution. In contrast to previous methods, our representation allows us to define a joint probabilistic learning objective that minimizes the dissimilarity between the two distributions. Our approach achieves state of the art results on the PASCAL VOC 2012 data set, outperforming the best baseline by $$4.2\%\ \text {mAP}^r_{0.5}$$ and $$4.8\%\ \text {mAP}^r_{0.75}$$ .

We present a visual localization framework based on novel deep attention aware features for autonomous driving that achieves centimeter level localization accuracy. Conventional approaches to the visual localization problem rely on handcrafted features or human-made objects on the road. They are known to be either prone to unstable matching caused by severe appearance or lighting changes, or too scarce to deliver constant and robust localization results in challenging scenarios. In this work, we seek to exploit the deep attention mechanism to search for salient, distinctive and stable features that are good for long-term matching in the scene through a novel end-to-end deep neural network. Furthermore, our learned feature descriptors are demonstrated to be competent to establish robust matches and therefore successfully estimate the optimal camera poses with high precision. We comprehensively validate the effectiveness of our method using a freshly collected dataset with high-quality ground truth trajectories and hardware synchronization between sensors. Results demonstrate that our method achieves a competitive localization accuracy when compared to the LiDAR-based localization solutions under various challenging circumstances, leading to a potential low-cost localization solution for autonomous driving.

We propose an end-to-end network for image generation from given structured-text that consists of the visual-relation layout module and stacking-GANs. Our visual-relation layout module uses relations among entities in the structured-text in two ways: comprehensive usage and individual usage. We comprehensively use all relations together to localize initial bounding-boxes (BBs) of all the entities. We use individual relation separately to predict from the initial BBs relation-units for all the relations. We then unify all the relation-units to produce the visual-relation layout, i.e., BBs for all the entities so that each of them uniquely corresponds to each entity while keeping its involved relations. Our visual-relation layout reflects the scene structure given in the input text. The stacking-GANs is the stack of three GANs conditioned on the visual-relation layout and the output of previous GAN, consistently capturing the scene structure. Our network realistically renders entities’ details while keeping the scene structure. Experimental results on two public datasets show the effectiveness of our method.

By adding human-imperceptible noise to clean images, the resultant adversarial examples can fool other unknown models. Features of a pixel extracted by deep neural networks (DNNs) are influenced by its surrounding regions, and different DNNs generally focus on different discriminative regions in recognition. Motivated by this, we propose a patch-wise iterative algorithm – a black-box attack towards mainstream normally trained and defense models, which differs from the existing attack methods manipulating pixel-wise noise. In this way, without sacrificing the performance of white-box attack, our adversarial examples can have strong transferability. Specifically, we introduce an amplification factor to the step size in each iteration, and one pixel’s overall gradient overflowing the $$\epsilon $$ -constraint is properly assigned to its surrounding regions by a project kernel. Our method can be generally integrated to any gradient-based attack methods. Compared with the current state-of-the-art attacks, we significantly improve the success rate by 9.2% for defense models and 3.7% for normally trained models on average. Our code is available at https://github.com/qilong-zhang/Patch-wise-iterative-attack

Feature interactions across space and scales underpin modern visual recognition systems because they introduce beneficial visual contexts. Conventionally, spatial contexts are passively hidden in the CNN’s increasing receptive fields or actively encoded by non-local convolution. Yet, the non-local spatial interactions are not across scales, and thus they fail to capture the non-local contexts of objects (or parts) residing in different scales. To this end, we propose a fully active feature interaction across both space and scales, called Feature Pyramid Transformer (FPT). It transforms any feature pyramid into another feature pyramid of the same size but with richer contexts, by using three specially designed transformers in self-level, top-down, and bottom-up interaction fashion. FPT serves as a generic visual backbone with fair computational overhead. We conduct extensive experiments in both instance-level (i.e., object detection and instance segmentation) and pixel-level segmentation tasks, using various backbones and head networks, and observe consistent improvement over all the baselines and the state-of-the-art methods (Code is open-sourced at https://github.com/ZHANGDONG-NJUST ).

Recently, end-to-end CNNs have presented remarkable performance for disparity estimation. But most of them are too heavy to resource-constrained devices, because of enormous parameters necessary for satisfactory results. To address the issue, we propose two compact stereo networks, MABNet and its light version MABNet_tiny. MABNet is based on a novel Multibranch Adjustable Bottleneck (MAB) module, which is less demanding on parameters and computation. In a MAB module, feature map is split into various parallel branches, where the depthwise separable convolutions with different dilation rates extract features with multiple receptive fields however at an affordable computational budget. Besides, the number of channels in each branch is adjustable independently to tradeoff computation and accuracy. On SceneFlow and KITTI datasets, our MABNet achieves competitive accuracy with fewer parameters of 1.65M. Especially, MABNet_tiny reduces the parameters 47K by cutting down the channels and layers in MABNet.

Person Re-identification (Re-ID) in crowed scenes is a challenging problem, where people are frequently partially occluded by objects and other people. However, few studies have provided flexible solutions to re-identifying people in an image containing a partial occlusion body part. In this paper, we propose a simple occlusion-aware approach to address the problem. The proposed method first leverages a fully convolutional network to generate spatial features. And then we design a combination of a pose-guided and mask-guided layer to generate saliency heatmap to further guide discriminative feature learning. More importantly, we propose a new matching approach, called Guided Adaptive Spatial Matching (GASM), which expects that each spatial feature in the query can find the most similar spatial features of a person in a gallery to match. Especially, We use the saliency heatmap to guide the adaptive spatial matching by assigning the foreground human parts with larger weights adaptively. The effectiveness of the proposed GASM is demonstrated on two occluded person datasets: Crowd REID (51.52%) and Occluded REID (80.25%) and three benchmark person datasets: Market1501 (95.31%), DukeMTMC-reID (88.12%) and MSMT17 (79.52%). Additionally, GASM achieves good performance on cross-domain person Re-ID. The code and models are available at: https://github.com/JDAI-CV/fast-reid/blob/master/projects/CrowdReID .

Most existing RGB-D saliency detection methods adopt symmetric two-stream architectures for learning discriminative RGB and depth representations. In fact, there is another level of ambiguity that is often overlooked: if RGB and depth data are necessary to fit into the same network. In this paper, we propose an asymmetric two-stream architecture taking account of the inherent differences between RGB and depth data for saliency detection. First, we design a flow ladder module (FLM) for the RGB stream to fully extract global and local information while maintaining the saliency details. This is achieved by constructing four detail-transfer branches, each of which preserves the detail information and receives global location information from representations of other vertical parallel branches in an evolutionary way. Second, we propose a novel depth attention module (DAM) to ensure depth features with high discriminative power in location and spatial structure being effectively utilized when combined with RGB features in challenging scenes. The depth features can also discriminatively guide the RGB features via our proposed DAM to precisely locate the salient objects. Extensive experiments demonstrate that our method achieves superior performance over 13 state-of-the-art RGB-D approaches on the 7 datasets. Our code will be publicly available.

The black-box nature of deep neural networks (DNNs) makes it impossible to understand why a particular output is produced, creating demand for “Explainable AI”. In this paper, we show that statistical fault localization (SFL) techniques from software engineering deliver high quality explanations of the outputs of DNNs, where we define an explanation as a minimal subset of features sufficient for making the same decision as for the original input. We present an algorithm and a tool called DeepCover, which synthesizes a ranking of the features of the inputs using SFL and constructs explanations for the decisions of the DNN based on this ranking. We compare explanations produced by DeepCover with those of the state-of-the-art tools gradcam, lime, shap, rise and extremal and show that explanations generated by DeepCover are consistently better across a broad set of experiments. On a benchmark set with known ground truth, DeepCover achieves $$76.7\%$$ accuracy, which is $$6\%$$ better than the second best extremal.

Building on recent progress at the intersection of combinatorial optimization and deep learning, we propose an end-to-end trainable architecture for deep graph matching that contains unmodified combinatorial solvers. Using the presence of heavily optimized combinatorial solvers together with some improvements in architecture design, we advance state-of-the-art on deep graph matching benchmarks for keypoint correspondence. In addition, we highlight the conceptual advantages of incorporating solvers into deep learning architectures, such as the possibility of post-processing with a strong multi-graph matching solver or the indifference to changes in the training setting. Finally, we propose two new challenging experimental setups.

We introduce a novel self-supervised learning approach to learn representations of videos that are responsive to changes in the motion dynamics. Our representations can be learned from data without human annotation and provide a substantial boost to the training of neural networks on small labeled data sets for tasks such as action recognition, which require to accurately distinguish the motion of objects. We promote an accurate learning of motion without human annotation by training a neural network to discriminate a video sequence from its temporally transformed versions. To learn to distinguish non-trivial motions, the design of the transformations is based on two principles: 1) To define clusters of motions based on time warps of different magnitude; 2) To ensure that the discrimination is feasible only by observing and analyzing as many image frames as possible. Thus, we introduce the following transformations: forward-backward playback, random frame skipping, and uniform frame skipping. Our experiments show that networks trained with the proposed method yield representations with improved transfer performance for action recognition on UCF101 and HMDB51.

In this paper, we look into the problem of estimating per-pixel depth maps from unconstrained RGB monocular night-time images which is a difficult task that has not been addressed adequately in the literature. The state-of-the-art day-time depth estimation methods fail miserably when tested with night-time images due to a large domain shift between them. The usual photometric losses used for training these networks may not work for night-time images due to the absence of uniform lighting which is commonly present in day-time images, making it a difficult problem to solve. We propose to solve this problem by posing it as a domain adaptation problem where a network trained with day-time images is adapted to work for night-time images. Specifically, an encoder is trained to generate features from night-time images that are indistinguishable from those obtained from day-time images by using a PatchGAN-based adversarial discriminative learning method. Unlike the existing methods that directly adapt depth prediction (network output), we propose to adapt feature maps obtained from the encoder network so that a pre-trained day-time depth decoder can be directly used for predicting depth from these adapted features. Hence, the resulting method is termed as “Adversarial Domain Feature Adaptation (ADFA)” and its efficacy is demonstrated through experimentation on the challenging Oxford night driving dataset. To the best of our knowledge, this work is a first of its kind to estimate depth from unconstrained night-time monocular RGB images that uses a completely unsupervised learning process. The modular encoder-decoder architecture for the proposed ADFA method allows us to use the encoder module as a feature extractor which can be used in many other applications. One such application is demonstrated where the features obtained from our adapted encoder network are shown to outperform other state-of-the-art methods in a visual place recognition problem, thereby, further establishing the usefulness and effectiveness of the proposed approach.

Connectionist Temporal Classification (CTC) is a training criterion designed for sequence labelling problems where the alignment between the inputs and the target labels is unknown. One of the key steps is to add a blank symbol to the target vocabulary. However, CTC tends to output spiky distributions since it prefers to output blank symbol most of the time. These spiky distributions show inferior alignments and the non-blank symbols are not learned sufficiently. To remedy this, we propose variational CTC (Var-CTC) to enhance the learning of non-blank symbols. The proposed Var-CTC converts the output distribution of vanilla CTC with hierarchy distribution. It first learns the approximated posterior distribution of blank to determine whether to output a specific non-blank symbol or not. Then it learns the alignment between non-blank symbols and input sequence. Experiments on scene text recognition and offline handwritten text recognition show Var-CTC achieves better alignments. Besides, with the enhanced learning of non-blank symbols, the confidence scores of model outputs are more discriminative. Compared with the vanilla CTC, the proposed Var-CTC can improve the recall performance by a large margin when the models maintain the same level of precision.

As an important computer vision task, 3d human pose estimation in a multi-camera, multi-person setting has received widespread attention and many interesting applications have been derived from it. Traditional approaches use a 3d pictorial structure model to handle this task. However, these models suffer from high computation costs and result in low accuracy in joint detection. Recently, especially since the introduction of Deep Neural Networks, one popular approach is to build a pipeline that involves three separate steps: (1) 2d skeleton detection in each camera view, (2) identification of matched 2d skeletons and (3) estimation of the 3d poses. Many existing works operate by feeding the 2d images and camera parameters through the three modules in a cascade fashion. However, all three operations can be highly correlated. For example, the 3d generation results may affect the results of detection in step 1, as does the matching algorithm in step 2. To address this phenomenon, we propose a novel end-to-end training scheme that brings the three separate modules into a single model. However, one outstanding problem of doing so is that the matching algorithm in step 2 appears to disjoint the pipeline. Therefore, we take our inspiration from the recent success in Capsule Networks, in which its Dynamic Routing step is also disjointed, but plays a crucial role in deciding how gradients are flowed from the upper to the lower layers. Similarly, a dynamic matching module in our work also decides the paths in which gradients flow from step 3 to step 1. Furthermore, as a large number of cameras are present, the existing matching algorithm either fails to deliver a robust performance or can be very inefficient. Thus, we additionally propose a novel matching algorithm that can match 2d poses from multiple views efficiently. The algorithm is robust and able to deal with situations of incomplete and false 2d detection as well.

In the real world, out-of-distribution samples, noise and distortions exist in test data. Existing deep networks developed for point cloud data analysis are prone to overfitting and a partial change in test data leads to unpredictable behaviour of the networks. In this paper, we propose a smart yet simple deep network for analysis of 3D models using ‘orderly disorder’ theory. Orderly disorder is a way of describing the complex structure of disorders within complex systems. Our method extracts the deep patterns inside a 3D object via creating a dynamic link to seek the most stable patterns and at once, throws away the unstable ones. Patterns are more robust to changes in data distribution, especially those that appear in the top layers. Features are extracted via an innovative cloning decomposition technique and then linked to each other to form stable complex patterns. Our model alleviates the vanishing-gradient problem, strengthens dynamic link propagation and substantially reduces the number of parameters. Extensive experiments on challenging benchmark datasets verify the superiority of our light network on the segmentation and classification tasks, especially in the presence of noise wherein our network’s performance drops less than 10% while the state-of-the-art networks fail to work.

Deep learning is emerging as a new paradigm for solving inverse imaging problems. However, the deep learning methods often lack the assurance of traditional physics-based methods due to the lack of physical information considerations in neural network training and deploying. The appropriate supervision and explicit calibration by the information of the physic model can enhance the neural network learning and its practical performance. In this paper, inspired by the geometry that data can be decomposed by two components from the null-space of the forward operator and the range space of its pseudo-inverse, we train neural networks to learn the two components and therefore learn the decomposition, i.e. we explicitly reformulate the neural network layers as learning range-nullspace decomposition functions with reference to the layer inputs, instead of learning unreferenced functions. We empirically show that the proposed framework demonstrates superior performance over recent deep residual learning, unrolled learning and nullspace learning on tasks including compressive sensing medical imaging and natural image super-resolution. Our code is available at https://github.com/edongdongchen/DDN .

We propose and study a method called FLOT that estimates scene flow on point clouds. We start the design of FLOT by noticing that scene flow estimation on point clouds reduces to estimating a permutation matrix in a perfect world. Inspired by recent works on graph matching, we build a method to find these correspondences by borrowing tools from optimal transport. Then, we relax the transport constraints to take into account real-world imperfections. The transport cost between two points is given by the pairwise similarity between deep features extracted by a neural network trained under full supervision using synthetic datasets. Our main finding is that FLOT can perform as well as the best existing methods on synthetic and real-world datasets while requiring much less parameters and without using multiscale analysis. Our second finding is that, on the training datasets considered, most of the performance can be explained by the learned transport cost. This yields a simpler method, FLOT $$_0$$ , which is obtained using a particular choice of optimal transport parameters and performs nearly as well as FLOT.

Affine correspondences (ACs) have been an active topic of research, namely for the recovery of surface normals. However, current solutions still suffer from the fact that even state-of-the-art affine feature detectors are inaccurate, and ACs are often contaminated by large levels of noise, yielding poor surface normals. This article provides new formulations for achieving epipolar geometry-consistent ACs, that, besides leading to linear solvers that are up to 30 $$\times $$ faster than the state-of-the-art alternatives, allow for a fast refinement scheme that significantly improves the quality of the noisy ACs. In addition, a tracker that automatically enforces the epipolar geometry is proposed, with experiments showing that it significantly outperforms competing methods in situations of low texture. This opens the way to application domains where the scenes are typically low textured, such as during arthroscopic procedures.

Existing techniques to encode spatial invariance within deep convolutional neural networks (CNNs) apply the same warping field to all the feature channels. This does not account for the fact that the individual feature channels can represent different semantic parts, which can undergo different spatial transformations w.r.t. a canonical configuration. To overcome this limitation, we introduce a learnable module, the volumetric transformer network (VTN), that predicts channel-wise warping fields so as to reconfigure intermediate CNN features spatially and channel-wisely. We design our VTN as an encoder-decoder network, with modules dedicated to letting the information flow across the feature channels, to account for the dependencies between the semantic parts. We further propose a loss function defined between the warped features of pairs of instances, which improves the localization ability of VTN. Our experiments show that VTN consistently boosts the features’ representation power and consequently the networks’ accuracy on fine-grained image recognition and instance-level image retrieval.

Panoramic 360 $$^{\circ }$$ images taken under unconstrained conditions present a significant challenge to current state-of-the-art recognition pipelines, since the assumption of a mostly upright camera is no longer valid. In this work, we investigate how to solve this problem by fusing purely geometric cues, such as apparent vanishing points, with learned semantic cues, such as the expectation that some visual elements (e.g. doors) have a natural upright position. We train a deep neural network to leverage these cues to segment the image-space endpoints of an imagined “vertical axis”, which is orthogonal to the ground plane of a scene, thus levelling the camera. We show that our segmentation-based strategy significantly increases performance, reducing errors by half, compared to the current state-of-the-art on two datasets of 360 $$^{\circ }$$ imagery. We also demonstrate the importance of 360 $$^{\circ }$$ camera levelling by analysing its impact on downstream tasks, finding that incorrect levelling severely degrades the performance of real-world computer vision pipelines.

Radon transform is a popular mathematical tool for shape analysis. However, it cannot handle affine deformation. Although its extended version, trace transform, allow us to construct affine invariants, they are less informative and computational expensive due to the loss of spatial relationship between trace lines and the extensive repeated calculation of transform. To address this issue, a novel line integral transform is proposed. We first use binding line pairs that have the desirable property of affine preserving as a reference frame to rewrite the diametrical dimension parameters of the lines in a relative manner which make them independent on affine transform. Along polar angle dimension of the line parameters, a moment-based normalization is then conducted to degrade the affine transform to similarity transform which can be easily normalized by Fourier transform. The proposed transform is not only invariant to affine transform, but also preserves the spatial relationship between line integrals which make it very informative. Another advantage of the proposed transform is that it is more efficient than the trace transform. Conducting it one time can allow us to achieve a 2D matrix of affine invariants. While conducting the trace transform once only generates a single feature and multiple trace transforms of different functionals are needed to derive more to make the descriptors informative. The effectiveness of the proposed transform has been validated on two types of standard shape test cases, affinely distorted contour shape dataset and region shape dataset, respectively.

Visual relationship detection is fundamental for holistic image understanding. However, the localization and classification of (subject, predicate, object) triplets remain challenging tasks, due to the combinatorial explosion of possible relationships, their long-tailed distribution in natural images, and an expensive annotation process.This paper introduces a novel weakly-supervised method for visual relationship detection that relies on minimal image-level predicate labels. A graph neural network is trained to classify predicates in images from a graph representation of detected objects, implicitly encoding an inductive bias for pairwise relations. We then frame relationship detection as the explanation of such a predicate classifier, i.e. we obtain a complete relation by recovering the subject and object of a predicted predicate.We present results comparable to recent fully- and weakly-supervised methods on three diverse and challenging datasets: HICO-DET for human-object interaction, Visual Relationship Detection for generic object-to-object relations, and UnRel for unusual triplets; demonstrating robustness to non-comprehensive annotations and good few-shot generalization.

Establishing dense semantic correspondences requires dealing with large geometric variations caused by the unconstrained setting of images. To address such severe matching ambiguities, we introduce a novel approach, called guided semantic flow, based on the key insight that sparse yet reliable matches can effectively capture non-rigid geometric variations, and these confident matches can guide adjacent pixels to have similar solution spaces, reducing the matching ambiguities significantly. We realize this idea with learning-based selection of confident matches from an initial set of all pairwise matching scores and their propagation by a new differentiable upsampling layer based on moving least square concept. We take advantage of the guidance from reliable matches to refine the matching hypotheses through Gaussian parametric model in the subsequent matching pipeline. With the proposed method, state-of-the-art performance is attained on several standard benchmarks for semantic correspondence.

Structure extraction from document images has been a long-standing research topic due to its high impact on a wide range of practical applications. In this paper, we share our findings on employing a hierarchical semantic segmentation network for this task of structure extraction. We propose a prior based deep hierarchical CNN network architecture that enables document structure extraction using very high resolution ( $$1800 \times 1000$$ ) images. We divide the document image into overlapping horizontal strips such that the network segments a strip and uses its prediction mask as prior for predicting the segmentation of the subsequent strip. We perform experiments establishing the effectiveness of our strip based network architecture through ablation methods and comparison with low-resolution variations. Further, to demonstrate our network’s capabilities, we train it on only one type of documents (Forms) and achieve state-of-the-art results over other general document datasets. We introduce our new human-annotated forms dataset and show that our method significantly outperforms different segmentation baselines on this dataset in extracting hierarchical structures. Our method is currently being used in Adobe’s AEM Forms for automated conversion of paper and PDF forms to modern HTML based forms.

Where people look when watching videos is believed to be heavily influenced by temporal patterns. In this work, we test this assumption by quantifying to which extent gaze on recent video saliency benchmarks can be predicted by a static baseline model. On the recent LEDOV dataset, we find that at least 75% of the explainable information as defined by a gold standard model can be explained using static features. Our baseline model “DeepGaze MR” even outperforms state-of-the-art video saliency models, despite deliberately ignoring all temporal patterns. Visual inspection of our static baseline’s failure cases shows that clear temporal effects on human gaze placement exist, but are both rare in the dataset and not captured by any of the recent video saliency models. To focus the development of video saliency models on better capturing temporal effects we construct a meta-dataset consisting of those examples requiring temporal information.

Self-driving cars need to understand 3D scenes efficiently and accurately in order to drive safely. Given the limited hardware resources, existing 3D perception models are not able to recognize small instances (e.g., pedestrians, cyclists) very well due to the low-resolution voxelization and aggressive downsampling. To this end, we propose Sparse Point-Voxel Convolution (SPVConv), a lightweight 3D module that equips the vanilla Sparse Convolution with the high-resolution point-based branch. With negligible overhead, this point-based branch is able to preserve the fine details even from large outdoor scenes. To explore the spectrum of efficient 3D models, we first define a flexible architecture design space based on SPVConv, and we then present 3D Neural Architecture Search (3D-NAS) to search the optimal network architecture over this diverse design space efficiently and effectively. Experimental results validate that the resulting SPVNAS model is fast and accurate: it outperforms the state-of-the-art MinkowskiNet by 3.3%, ranking 1 $$^\mathbf{st}$$ on the competitive SemanticKITTI leaderboard $$^\star $$ . It also achieves 8–23 $$\times $$ computation reduction and 3 $$\times $$ measured speedup over MinkowskiNet and KPConv with higher accuracy. Finally, we transfer our method to 3D object detection, and it achieves consistent improvements over the one-stage detection baseline on KITTI.

Existing connectivity-oriented performance measures rank road delineation algorithms inconsistently, which makes it difficult to decide which one is best for a given application. We show that these inconsistencies stem from design flaws that make the metrics insensitive to whole classes of errors. This insensitivity is undesirable in metrics intended for capturing overall general quality of road reconstructions. In particular, the scores do not reflect the time needed for a human to fix the errors, because each one has to be fixed individually. To provide more reliable evaluation, we design three new metrics that are sensitive to all classes of errors. This sensitivity makes them more consistent even though they use very different approaches to comparing ground-truth and reconstructed road networks. We use both synthetic and real data to demonstrate this and advocate the use of these corrected metrics as a tool to gauge future progress.

Continual Learning (CL) aims to develop agents emulating the human ability to sequentially learn new tasks while being able to retain knowledge obtained from past experiences. In this paper, we introduce the novel problem of Memory-Constrained Online Continual Learning (MC-OCL) which imposes strict constraints on the memory overhead that a possible algorithm can use to avoid catastrophic forgetting. As most, if not all, previous CL methods violate these constraints, we propose an algorithmic solution to MC-OCL: Batch-level Distillation (BLD), a regularization-based CL approach, which effectively balances stability and plasticity in order to learn from data streams, while preserving the ability to solve old tasks through distillation. Our extensive experimental evaluation, conducted on three publicly available benchmarks, empirically demonstrates that our approach successfully addresses the MC-OCL problem and achieves comparable accuracy to prior distillation methods requiring higher memory overhead (Code available at https://github.com/DonkeyShot21/batch-level-distillation ).

While unsupervised domain adaptation methods based on deep architectures have achieved remarkable success in many computer vision tasks, they rely on a strong assumption, i.e. labeled source data must be available. In this work we overcome this assumption and we address the problem of transferring knowledge from a source to a target domain when both source and target data have no annotations. Inspired by recent works on deep clustering, our approach leverages information from data gathered from multiple source domains to build a domain-agnostic clustering model which is then refined at inference time when target data become available. Specifically, at training time we propose to optimize a novel information-theoretic loss which, coupled with domain-alignment layers, ensures that our model learns to correctly discover semantic labels while discarding domain-specific features. Importantly, our architecture design ensures that at inference time the resulting source model can be effectively adapted to the target domain without having access to source data, thanks to feature alignment and self-supervision. We evaluate the proposed approach in a variety of settings (Code available at https://github.com/willi-menapace/acids-clustering-domain-shift ), considering several domain adaptation benchmarks and we show that our method is able to automatically discover relevant semantic information even in presence of few target samples and yields state-of-the-art results on multiple domain adaptation benchmarks.

Deep neural networks are vulnerable to adversarial attacks. Though various attempts have been made, it is still largely open to fully understand the existence of adversarial samples and thereby develop effective defense strategies. In this paper, we present a new perspective, namely gradient leaking hypothesis, to understand the existence of adversarial examples and to further motivate effective defense strategies. Specifically, we consider the low dimensional manifold structure of natural images, and empirically verify that the leakage of the gradient (w.r.t input) along the (approximately) perpendicular direction to the tangent space of data manifold is a reason for the vulnerability over adversarial attacks. Based on our investigation, we further present a new robust learning algorithm which encourages a larger gradient component in the tangent space of data manifold, suppressing the gradient leaking phenomenon consequently. Experiments on various tasks demonstrate the effectiveness of our algorithm despite its simplicity.

In this work we review the coarse-to-fine spatial feature pyramid concept, which is used in state-of-the-art optical flow estimation networks to make exploration of the pixel flow search space computationally tractable and efficient. Within an individual pyramid level, we improve the cost volume construction process by departing from a warping- to a sampling-based strategy, which avoids ghosting and hence enables us to better preserve fine flow details. We further amplify the positive effects through a level-specific, loss max-pooling strategy that adaptively shifts the focus of the learning process on under-performing predictions. Our second contribution revises the gradient flow across pyramid levels. The typical operations performed at each pyramid level can lead to noisy, or even contradicting gradients across levels. We show and discuss how properly blocking some of these gradient components leads to improved convergence and ultimately better performance. Finally, we introduce a distillation concept to counteract the issue of catastrophic forgetting during finetuning and thus preserving knowledge over models sequentially trained on multiple datasets. Our findings are conceptually simple and easy to implement, yet result in compelling improvements on relevant error measures that we demonstrate via exhaustive ablations on datasets like Flying Chairs2, Flying Things, Sintel and KITTI. We establish new state-of-the-art results on the challenging Sintel and KITTI 2012 test datasets, and even show the portability of our findings to different optical flow and depth from stereo approaches.