Computer Vision – ECCV 2020

The connection between visual input and tactile sensing is critical for object manipulation tasks such as grasping and pushing. In this work, we introduce the challenging task of estimating a set of tactile physical properties from visual information. We aim to build a model that learns the complex mapping between visual information and tactile physical properties. We construct a first of its kind image-tactile dataset with over 400 multiview image sequences and the corresponding tactile properties. A total of fifteen tactile physical properties across categories including friction, compliance, adhesion, texture, and thermal conductance are measured and then estimated by our models. We develop a cross-modal framework comprised of an adversarial objective and a novel visuo-tactile joint classification loss. Additionally, we introduce a neural architecture search framework capable of selecting optimal combinations of viewing angles for estimating a given physical property.

Fitting a matrix of a given rank to data in a least squares sense can be done very effectively using 2nd order methods such as Levenberg-Marquardt by explicitly optimizing over a bilinear parameterization of the matrix. In contrast, when applying more general singular value penalties, such as weighted nuclear norm priors, direct optimization over the elements of the matrix is typically used. Due to non-differentiability of the resulting objective function, first order sub-gradient or splitting methods are predominantly used. While these offer rapid iterations it is well known that they become inefficient near the minimum due to zig-zagging and in practice one is therefore often forced to settle for an approximate solution.In this paper we show that more accurate results can in many cases be achieved with 2nd order methods. Our main result shows how to construct bilinear formulations, for a general class of regularizers including weighted nuclear norm penalties, that are provably equivalent to the original problems. With these formulations the regularizing function becomes twice differentiable and 2nd order methods can be applied. We show experimentally, on a number of structure from motion problems, that our approach outperforms state-of-the-art methods.

Video inpainting is an important technique for a wide variety of applications from video content editing to video restoration. Early approaches follow image inpainting paradigms, but are challenged by complex camera motion and non-rigid deformations. To address these challenges flow-guided propagation techniques have been proposed. However, computation of flow is non-trivial for unobserved regions and propagation across a whole video sequence is computationally demanding. In contrast, in this paper, we propose a video inpainting algorithm based on proposals: we use 3D convolutions to obtain an initial inpainting estimate which is subsequently refined by fusing a generated set of proposals. Different from existing approaches for video inpainting, and inspired by well-explored mechanisms for object detection, we argue that proposals provide a rich source of information that permits combining similarly looking patches that may be spatially and temporally far from the region to be inpainted. We validate the effectiveness of our method on the challenging YouTube VOS and DAVIS datasets using different settings and demonstrate results outperforming state-of-the-art on standard metrics.

Domain Adaptation as an important tool aims to explore a generalized model trained on well-annotated source knowledge to address learning issue on target domain with insufficient or even no annotation. Current approaches typically incorporate data from source and target domains for training stage to deal with domain shift. However, most domain adaptation tasks generally suffer from the problem that measuring the domain shift tends to be impossible when target data is inaccessible. In this paper, we propose a novel algorithm, Hybrid Generative Network (HGNet) for Zero-shot Domain Adaptation, which embeds an adaptive feature separation (AFS) module into generative architecture. Specifically, AFS module can adaptively distinguish classification-relevant features from classification-irrelevant ones to learn domain-invariant and discriminative representations when task-relevant target instances are invisible. To learn high-quality feature representation, we also develop hybrid generative strategy to ensure the uniqueness of feature separation and completeness of semantic information. Extensive experimental results on several benchmarks illustrate that our method achieves more promising results than state-of-the-art approaches.

Rain is a common natural phenomenon. Taking images in the rain however often results in degraded quality of images, thus compromises the performance of many computer vision systems. Most existing de-rain algorithms use only one single input image and aim to recover a clean image. Few work has exploited stereo images. Moreover, even for single image based monocular deraining, many current methods fail to complete the task satisfactorily because they mostly rely on per pixel loss functions and ignore semantic information. In this paper, we present a Paired Rain Removal Network (PRRNet), which exploits both stereo images and semantic information. Specifically, we develop a Semantic-Aware Deraining Module (SADM) which solves both tasks of semantic segmentation and deraining of scenes, and a Semantic-Fusion Network (SFNet) and a View-Fusion Network (VFNet) which fuse semantic information and multi-view information respectively. We also propose new stereo based rainy datasets for benchmarking. Experiments on both monocular and the newly proposed stereo rainy datasets demonstrate that the proposed method achieves the state-of-the-art performance.

Deep neural networks have achieved state-of-the art performance on various computer vision tasks. However, their deployment on resource-constrained devices has been hindered due to their high computational and storage complexity. While various complexity reduction techniques, such as lightweight network architecture design and parameter quantization, have been successful in reducing the cost of implementing these networks, these methods have often been considered orthogonal. In reality, existing quantization techniques fail to replicate their success on lightweight architectures such as MobileNet. To this end, we present a novel fully differentiable non-uniform quantizer that can be seamlessly mapped onto efficient ternary-based dot product engines. We conduct comprehensive experiments on CIFAR-10, ImageNet, and Visual Wake Words datasets. The proposed quantizer (DBQ) successfully tackles the daunting task of aggressively quantizing lightweight networks such as MobileNetV1, MobileNetV2, and ShuffleNetV2. DBQ achieves state-of-the art results with minimal training overhead and provides the best (pareto-optimal) accuracy-complexity trade-off.

Recent advances in high refresh rate displays as well as the increased interest in high rate of slow motion and frame up-conversion fuel the demand for efficient and cost-effective multi-frame video interpolation solutions. To that regard, inserting multiple frames between consecutive video frames are of paramount importance for the consumer electronics industry. State-of-the-art methods are iterative solutions interpolating one frame at the time. They introduce temporal inconsistencies and clearly noticeable visual artifacts.Departing from the state-of-the-art, this work introduces a true multi-frame interpolator. It utilizes a pyramidal style network in the temporal domain to complete the multi-frame interpolation task in one-shot. A novel flow estimation procedure using a relaxed loss function, and an advanced, cubic-based, motion model is also used to further boost interpolation accuracy when complex motion segments are encountered. Results on the Adobe240 dataset show that the proposed method generates visually pleasing, temporally consistent frames, outperforms the current best off-the-shelf method by 1.57 dB in PSNR with 8 times smaller model and 7.7 times faster. The proposed method can be easily extended to interpolate a large number of new frames while remaining efficient because of the one-shot mechanism.

Recent progress on few-shot learning largely relies on annotated data for meta-learning: base classes sampled from the same domain as the novel classes. However, in many applications, collecting data for meta-learning is infeasible or impossible. This leads to the cross-domain few-shot learning problem, where there is a large shift between base and novel class domains. While investigations of the cross-domain few-shot scenario exist, these works are limited to natural images that still contain a high degree of visual similarity. No work yet exists that examines few-shot learning across different imaging methods seen in real world scenarios, such as aerial and medical imaging. In this paper, we propose the Broader Study of Cross-Domain Few-Shot Learning (BSCD-FSL) benchmark, consisting of image data from a diverse assortment of image acquisition methods. This includes natural images, such as crop disease images, but additionally those that present with an increasing dissimilarity to natural images, such as satellite images, dermatology images, and radiology images. Extensive experiments on the proposed benchmark are performed to evaluate state-of-art meta-learning approaches, transfer learning approaches, and newer methods for cross-domain few-shot learning. The results demonstrate that state-of-art meta-learning methods are surprisingly outperformed by earlier meta-learning approaches, and all meta-learning methods underperform in relation to simple fine-tuning by 12.8% average accuracy. In some cases, meta-learning even underperforms networks with random weights. Performance gains previously observed with methods specialized for cross-domain few-shot learning vanish in this more challenging benchmark. Finally, accuracy of all methods tend to correlate with dataset similarity to natural images, verifying the value of the benchmark to better represent the diversity of data seen in practice and guiding future research. Code for the experiments in this work can be found at https://github.com/IBM/cdfsl-benchmark .

Data poisoning attacks on machine learning models have attracted much recent attention, wherein poisoning samples are injected at the training phase to achieve adversarial goals at test time. Although existing poisoning techniques prove to be effective in various scenarios, they rely on certain assumptions on the adversary knowledge and capability to ensure efficacy, which may be unrealistic in practice. This paper presents a new, practical targeted poisoning attack method on neural networks in vision domain, namely BlackCard. BlackCard possesses a set of critical properties for ensuring attacking efficacy in practice, which has never been simultaneously achieved by any existing work, including knowledge-oblivious, clean-label, and clean-test. Importantly, we show that the effectiveness of BlackCard can be intuitively guaranteed by a set of analytical reasoning and observations, through exploiting an essential characteristic of gradient-descent optimization which is pervasively adopted in DNN models. We evaluate the efficacy of BlackCard for generating targeted poisoning attacks via extensive experiments using various datasets and DNN models. Results show that BlackCard is effective with a rather high success rate while preserving all the claimed properties.

Person re-identification (ReID) remains a challenging task in many real-word video analytics and surveillance applications, even though state-of-the-art accuracy has improved considerably with the advent of deep learning (DL) models trained on large image datasets. Given the shift in distributions that typically occurs between video data captured from the source and target domains, and absence of labeled data from the target domain, it is difficult to adapt a DL model for accurate recognition of target data. DL models for unsupervised domain adaptation (UDA) are commonly designed in the feature representation space. We argue that for pair-wise matchers that rely on metric learning, e.g., Siamese networks for person ReID, the UDA objective should consist in aligning pair-wise dissimilarity between domains, rather than aligning feature representations. Moreover, dissimilarity representations are more suitable for designing open-set ReID systems, where identities differ in the source and target domains. In this paper, we propose a novel Dissimilarity-based Maximum Mean Discrepancy (D-MMD) loss for aligning pair-wise distances that can be optimized via gradient descent using relatively small batch sizes. From a person ReID perspective, the evaluation of D-MMD loss is straightforward since the tracklet information (provided by a person tracker) allows to label a distance vector as being either within-class (within-tracklet) or between-class (between-tracklet). This allows approximating the underlying distribution of target pair-wise distances for D-MMD loss optimization, and accordingly align source and target distance distributions. Empirical results with three challenging benchmark datasets show that the proposed D-MMD loss decreases as source and domain distributions become more similar. Extensive experimental evaluation also indicates that UDA methods that rely on the D-MMD loss can significantly outperform baseline and state-of-the-art UDA methods for person ReID. The dissimilarity space transformation allows to design reliable pair-wise matchers, without the common requirement for data augmentation and/or complex networks. Code is available on GitHub link: https://github.com/djidje/D-MMD .

In this work, we propose a method for training distributed GAN with sequential temporary discriminators. Our proposed method tackles the challenge of training GAN in the federated learning manner: How to update the generator with a flow of temporary discriminators? We apply our proposed method to learn a self-adaptive generator with a series of local discriminators from multiple data centers. We show our design of loss function indeed learns the correct distribution with provable guarantees. The empirical experiments show that our approach is capable of generating synthetic data which is practical for real-world applications such as training a segmentation model. Our TDGAN Code is available at: https://github.com/huiqu18/TDGAN-PyTorch .

We propose a system comprised of fixed-topology neural networks having partially frozen weights, named SemifreddoNets. SemifreddoNets work as fully-pipelined hardware blocks that are optimized to have an efficient hardware implementation. Those blocks freeze a certain portion of the parameters at every layer and replace the corresponding multipliers with fixed scalers. Fixing the weights reduces the silicon area, logic delay, and memory requirements, leading to significant savings in cost and power consumption. Unlike traditional layer-wise freezing approaches, SemifreddoNets make a profitable trade between the cost and flexibility by having some of the weights configurable at different scales and levels of abstraction in the model. Although fixing the topology and some of the weights somewhat limits the flexibility, we argue that the efficiency benefits of this strategy outweigh the advantages of a fully configurable model for many use cases. Furthermore, our system uses repeatable blocks, therefore it has the flexibility to adjust model complexity without requiring any hardware change. The hardware implementation of SemifreddoNets provides up to an order of magnitude reduction in silicon area and power consumption as compared to their equivalent implementation on a general-purpose accelerator.

The fact that deep neural networks are susceptible to crafted perturbations severely impacts the use of deep learning in certain domains of application. Among many developed defense models against such attacks, adversarial training emerges as the most successful method that consistently resists a wide range of attacks. In this work, based on an observation from a previous study that the representations of a clean data example and its adversarial examples become more divergent in higher layers of a deep neural net, we propose the Adversary Divergence Reduction Network which enforces local/global compactness and the clustering assumption over an intermediate layer of a deep neural network. We conduct comprehensive experiments to understand the isolating behavior of each component (i.e., local/global compactness and the clustering assumption) and compare our proposed model with state-of-the-art adversarial training methods. The experimental results demonstrate that augmenting adversarial training with our proposed components can further improve the robustness of the network, leading to higher unperturbed and adversarial predictive performances.

Most state-of-the-art approaches to road extraction from aerial images rely on a CNN trained to label road pixels as foreground and remainder of the image as background. The CNN is usually trained by minimizing pixel-wise losses, which is less than ideal to produce binary masks that preserve the road network’s global connectivity.To address this issue, we introduce an Adversarial Learning (AL) strategy tailored for our purposes. A naive one would treat the segmentation network as a generator and would feed its output along with ground-truth segmentations to a discriminator. It would then train the generator and discriminator jointly. We will show that this is not enough because it does not capture the fact that most errors are local and need to be treated as such. Instead, we use a more sophisticated discriminator that returns a label pyramid describing what portions of the road network are correct at several different scales.This discriminator and the structured labels it returns are what gives our approach its edge and we will show that it outperforms state-of-the-art ones on the challenging RoadTracer dataset.

Important applications such as mobile computing require reducing the computational costs of neural network inference. Ideally, applications would specify their preferred tradeoff between accuracy and speed, and the network would optimize this end-to-end, using classification error to remove parts of the network. Increasing speed can be done either during training – e.g., pruning filters – or during inference – e.g., conditionally executing a subset of the layers. We propose a single end-to-end framework that can improve inference efficiency in both settings. We use a combination of batch activation loss and classification loss, and Gumbel reparameterization to learn network structure. We train end-to-end, and the same technique supports pruning as well as conditional computation. We obtain promising experimental results for ImageNet classification with ResNet (45–52% less computation).

While supervised techniques in re-identification are extremely effective, the need for large amounts of annotations makes them impractical for large camera networks. One-shot re-identification, which uses a singular labeled tracklet for each identity along with a pool of unlabeled tracklets, is a potential candidate towards reducing this labeling effort. Current one-shot re-identification methods function by modeling the inter-relationships amongst the labeled and the unlabeled data, but fail to fully exploit such relationships that exist within the pool of unlabeled data itself. In this paper, we propose a new framework named Temporal Consistency Progressive Learning, which uses temporal coherence as a novel self-supervised auxiliary task in the one-shot learning paradigm to capture such relationships amongst the unlabeled tracklets. Optimizing two new losses, which enforce consistency on a local and global scale, our framework can learn richer and more discriminative representations. Extensive experiments on two challenging video re-identification datasets - MARS and DukeMTMC-VideoReID - demonstrate that our proposed method is able to estimate the true labels of the unlabeled data more accurately by up to $$8\%$$ 8 % , and obtain significantly better re-identification performance compared to the existing state-of-the-art techniques .

As machine learning models and dataset escalate in scales rapidly, the huge memory footprint impedes efficient training. Reversible operators can reduce memory consumption by discarding intermediate feature maps in forward computations and recover them via their inverse functions in the backward propagation. They save memory at the cost of computation overhead. However, current implementations of reversible layers mainly focus on saving memory usage with computation overhead neglected. In this work, we formulate the decision problem for reversible operators with training time as the objective function and memory usage as the constraint. By solving this problem, we can maximize the training throughput for reversible neural architectures. Our proposed framework fully automates this decision process, empowering researchers to develop and train reversible neural networks more efficiently.

We propose a weakly supervised approach to semantic segmentation using bounding box annotations. Bounding boxes are treated as noisy labels for the foreground objects. We predict a per-class attention map that saliently guides the per-pixel cross entropy loss to focus on foreground pixels and refines the segmentation boundaries. This avoids propagating erroneous gradients due to incorrect foreground labels on the background. Additionally, we learn pixel embeddings to simultaneously optimize for high intra-class feature affinity while increasing discrimination between features across different classes. Our method, Box2Seg, achieves state-of-the-art segmentation accuracy on PASCAL VOC 2012 by significantly improving the mIOU metric by $$2.1\%$$ 2.1 % compared to previous weakly supervised approaches. Our weakly supervised approach is comparable to the recent fully supervised methods when fine-tuned with limited amount of pixel-level annotations. Qualitative results and ablation studies show the benefit of different loss terms on the overall performance.

A 3D imaging and mapping system that can handle both multiple-viewers and dynamic-objects is attractive for many applications. We propose a freeform structured light system that does not rigidly constrain camera(s) to the projector. By introducing an optimized phase-coded aperture in the projector, we transform the projector pattern to encode depth in its defocus robustly; this allows a camera to estimate depth, in projector coordinates, using local information. Additionally, we project a Kronecker-multiplexed pattern that provides global context to establish correspondence between camera and projector pixels. Together with aperture coding and projected pattern, the projector offers a unique 3D labeling for every location of the scene. The projected pattern can be observed in part or full by any camera, to reconstruct both the 3D map of the scene and the camera pose in the projector coordinates. This system is optimized using a fully differentiable rendering model and a CNN-based reconstruction. We build a prototype and demonstrate high-quality 3D reconstruction with an unconstrained camera, for both dynamic scenes and multi-camera systems.

We tackle the convolution neural networks (CNNs) backdoor detection problem by proposing a new representation called one-pixel signature. Our task is to detect/classify if a CNN model has been maliciously inserted with an unknown Trojan trigger or not. We design the one-pixel signature representation to reveal the characteristics of both clean and backdoored CNN models. Here, each CNN model is associated with a signature that is created by generating, pixel-by-pixel, an adversarial value that is the result of the largest change to the class prediction. The one-pixel signature is agnostic to the design choice of CNN architectures, and how they were trained. It can be computed efficiently for a black-box CNN model without accessing the network parameters. Our proposed one-pixel signature demonstrates a substantial improvement (by around 30% in the absolute detection accuracy) over the existing competing methods for backdoored CNN detection/classification. One-pixel signature is a general representation that can be used to characterize CNN models beyond backdoor detection.

We propose a novel adaptive transfer learning framework, learning to transfer learn (L2TL), to improve performance on a target dataset by careful extraction of the related information from a source dataset. Our framework considers cooperative optimization of shared weights between models for source and target tasks, and adjusts the constituent loss weights adaptively. The adaptation of the weights is based on a reinforcement learning (RL) selection policy, guided with a performance metric on the target validation set. We demonstrate that L2TL outperforms fine-tuning baselines and other adaptive transfer learning methods on eight datasets. In the regimes of small-scale target datasets and significant label mismatch between source and target datasets, L2TL shows particularly large benefits.

Sharing food has become very popular with the development of social media. For many real-world applications, people are keen to know the underlying recipes of a food item. In this paper, we are interested in automatically generating cooking instructions for food. We investigate an open research task of generating cooking instructions based on only food images and ingredients, which is similar to the image captioning task. However, compared with image captioning datasets, the target recipes are long-length paragraphs and do not have annotations on structure information. To address the above limitations, we propose a novel framework of Structure-aware Generation Network (SGN) to tackle the food recipe generation task. Our approach brings together several novel ideas in a systematic framework: (1) exploiting an unsupervised learning approach to obtain the sentence-level tree structure labels before training; (2) generating trees of target recipes from images with the supervision of tree structure labels learned from (1); and (3) integrating the inferred tree structures with the recipe generation procedure. Our proposed model can produce high-quality and coherent recipes, and achieve the state-of-the-art performance on the benchmark Recipe1M dataset.

Deep metric learning (DML) has received much attention in deep learning due to its wide applications in computer vision. Previous studies have focused on designing complicated losses and hard example mining methods, which are mostly heuristic and lack of theoretical understanding. In this paper, we cast DML as a simple pairwise binary classification problem that classifies a pair of examples as similar or dissimilar. It identifies the most critical issue in this problem—imbalanced data pairs. To tackle this issue, we propose a simple and effective framework to sample pairs in a batch of data for updating the model. The key to this framework is to define a robust loss for all pairs over a mini-batch of data, which is formulated by distributionally robust optimization. The flexibility in constructing the uncertainty decision set of the dual variable allows us to recover state-of-the-art complicated losses and also to induce novel variants. Empirical studies on several benchmark data sets demonstrate that our simple and effective method outperforms the state-of-the-art results.

Remote heart rate estimation is the measurement of heart rate without any physical contact with the subject and is accomplished using remote photoplethysmography (rPPG) in this work. rPPG signals are usually collected using a video camera with a limitation of being sensitive to multiple contributing factors, e.g. variation in skin tone, lighting condition and facial structure. End-to-end supervised learning approach performs well when training data is abundant, covering a distribution that doesn’t deviate too much from the distribution of testing data or during deployment. To cope with the unforeseeable distributional changes during deployment, we propose a transductive meta-learner that takes unlabeled samples during testing (deployment) for a self-supervised weight adjustment (also known as transductive inference), providing fast adaptation to the distributional changes. Using this approach, we achieve state-of-the-art performance on MAHNOB-HCI and UBFC-rPPG.

In this work, we address the problem of synthesizing human actions from novel views. Given an input video of an actor performing some action, we aim to synthesize a video with the same action performed from a novel view with the help of an appearance prior. We propose an end-to-end deep network to solve this problem. The proposed network utilizes the change in viewpoint to transform the action from the input view to the novel view in feature space. The transformed action is integrated with the target appearance using the proposed recurrent transformer network, which provides a transformed appearance for each time-step in the action sequence. The recurrent transformer network utilize action key-points which are determined in an unsupervised approach using the encoded action features. We also propose a hierarchical structure for the recurrent transformation which further improves the performance. We demonstrate the effectiveness of the proposed method through extensive experiments conducted on a large-scale multi-view action recognition NTU-RGB+D dataset. In addition, we show that the proposed method can transform the action to a novel viewpoint with an entirely different scene or actor. The code is publicly available at https://github.com/schatzkara/cross-view-video .

In this work, we address the problem of action recognition in a multi-view environment. Most of the existing approaches utilize pose information for multi-view action recognition. We focus on RGB modality instead and propose an unsupervised representation learning framework, which encodes the scene dynamics in videos captured from multiple viewpoints via predicting actions from unseen views. The framework takes multiple short video clips from different viewpoints and time as input and learns an holistic internal representation which is used to predict a video clip from an unseen viewpoint and time. The ability of the proposed network to render unseen video frames enables it to learn a meaningful and robust representation of the scene dynamics. We evaluate the effectiveness of the learned representation for multi-view video action recognition in a supervised approach. We observe a significant improvement in the performance with RGB modality on NTU-RGB+D dataset, which is the largest dataset for multi-view action recognition. The proposed framework also achieves state-of-the-art results with depth modality, which validates the generalization capability of the approach to other data modalities. The code is publicly available at https://github.com/svyas23/cross-view-action .

In this paper, we introduce a novel network, called discriminative feature network (DFNet), to address the unsupervised video object segmentation task. To capture the inherent correlation among video frames, we learn discriminative features (D-features) from the input images that reveal feature distribution from a global perspective. The D-features are then used to establish correspondence with all features of test image under conditional random field (CRF) formulation, which is leveraged to enforce consistency between pixels. The experiments verify that DFNet outperforms state-of-the-art methods by a large margin with a mean IoU score of 83.4% and ranks first on the DAVIS-2016 leaderboard while using much fewer parameters and achieving much more efficient performance in the inference phase. We further evaluate DFNet on the FBMS dataset and the video saliency dataset ViSal, reaching a new state-of-the-art. To further demonstrate the generalizability of our framework, DFNet is also applied to the image object co-segmentation task. We perform experiments on a challenging dataset PASCAL-VOC and observe the superiority of DFNet. The thorough experiments verify that DFNet is able to capture and mine the underlying relations of images and discover the common foreground objects.

We propose advances that address two key challenges in future trajectory prediction: (i) multimodality in both training data and predictions and (ii) constant time inference regardless of number of agents. Existing trajectory predictions are fundamentally limited by lack of diversity in training data, which is difficult to acquire with sufficient coverage of possible modes. Our first contribution is an automatic method to simulate diverse trajectories in the top-view. It uses pre-existing datasets and maps as initialization, mines existing trajectories to represent realistic driving behaviors and uses a multi-agent vehicle dynamics simulator to generate diverse new trajectories that cover various modes and are consistent with scene layout constraints. Our second contribution is a novel method that generates diverse predictions while accounting for scene semantics and multi-agent interactions, with constant-time inference independent of the number of agents. We propose a convLSTM with novel state pooling operations and losses to predict scene-consistent states of multiple agents in a single forward pass, along with a CVAE for diversity. We validate our proposed multi-agent trajectory prediction approach by training and testing on the proposed simulated dataset and existing real datasets of traffic scenes. In both cases, our approach outperforms SOTA methods by a large margin, highlighting the benefits of both our diverse dataset simulation and constant-time diverse trajectory prediction methods.

Unsupervised domain adaptation enables to alleviate the need for pixel-wise annotation in the semantic segmentation. One of the most common strategies is to translate images from the source domain to the target domain and then align their marginal distributions in the feature space using adversarial learning. However, source-to-target translation enlarges the bias in translated images and introduces extra computations, owing to the dominant data size of the source domain. Furthermore, consistency of the joint distribution in source and target domains cannot be guaranteed through global feature alignment. Here, we present an innovative framework, designed to mitigate the image translation bias and align cross-domain features with the same category. This is achieved by 1) performing the target-to-source translation and 2) reconstructing both source and target images from their predicted labels. Extensive experiments on adapting from synthetic to real urban scene understanding demonstrate that our framework competes favorably against existing state-of-the-art methods.

Large-scale point cloud semantic understanding is an important problem in self-driving cars and autonomous robotics navigation. However, such problem involves many challenges, such as i) critical road objects (e.g., pedestrians, barriers) with diverse and varying input shapes; ii) distributed contextual information across large spatial range; iii) efficient inference time. Failing to deal with such challenges may weaken the mission-critical performance of self-driving car, e.g, LiDAR road objects perception. In this work, we propose a novel neural network model called Attention-based Dynamic Convolution Network with Self-Attention Global Contexts(ADConvnet-SAGC), which i) applies attention mechanism to adaptively focus on the most task-related neighboring points for learning the point features of 3D objects, especially for small objects with diverse shapes; ii) applies self-attention module for efficiently capturing long-range distributed contexts from the input; iii) a more reasonable and compact architecture for efficient inference. Extensive experiments on point cloud semantic segmentation validate the effectiveness of the proposed ADConvnet-SAGC model and show significant improvements over state-of-the-art methods.

Interpretability is an emerging area of research in trustworthy machine learning. Safe deployment of machine learning system mandates that the prediction and its explanation be reliable and robust. Recently, it has been shown that the explanations could be manipulated easily by adding visually imperceptible perturbations to the input while keeping the model’s prediction intact. In this work, we study the problem of attributional robustness (i.e. models having robust explanations) by showing an upper bound for attributional vulnerability in terms of spatial correlation between the input image and its explanation map. We propose a training methodology that learns robust features by minimizing this upper bound using soft-margin triplet loss. Our methodology of robust attribution training (ART) achieves the new state-of-the-art attributional robustness measure by a margin of $$\approx $$ ≈ 6–18 $$\%$$ % on several standard datasets, ie. SVHN, CIFAR-10 and GTSRB. We further show the utility of the proposed robust training technique (ART) in the downstream task of weakly supervised object localization by achieving the new state-of-the-art performance on CUB-200 dataset.

Event cameras are paradigm-shifting novel sensors that report asynchronous, per-pixel brightness changes called ‘events’ with unparalleled low latency. This makes them ideal for high speed, high dynamic range scenes where conventional cameras would fail. Recent work has demonstrated impressive results using Convolutional Neural Networks (CNNs) for video reconstruction and optic flow with events. We present strategies for improving training data for event based CNNs that result in 20–40% boost in performance of existing state-of-the-art (SOTA) video reconstruction networks retrained with our method, and up to 15% for optic flow networks. A challenge in evaluating event based video reconstruction is lack of quality ground truth images in existing datasets. To address this, we present a new High Quality Frames (HQF) dataset, containing events and ground truth frames from a DAVIS240C that are well-exposed and minimally motion-blurred. We evaluate our method on HQF + several existing major event camera datasets.

Unlike most existing works that define room layout on a 2D image, we model the layout in 3D as a configuration of the camera and the room. Our spatial geometric representation with only seven variables is more concise but effective, and more importantly enables direct 3D reasoning, e.g. how the camera is positioned relative to the room. This is particularly valuable in applications such as indoor robot navigation. We formulate the problem as a Markov decision process, in which the layout is incrementally adjusted based on the difference between the current layout and the target image, and the policy is learned via deep reinforcement learning. Our framework is end-to-end trainable, requiring no extra optimization, and achieves competitive performance on two challenging room layout datasets.

Much research on object detection focuses on building better model architectures and detection algorithms. Changing the model architecture, however, comes at the cost of adding more complexity to inference, making models slower. Data augmentation, on the other hand, doesn’t add any inference complexity, but is insufficiently studied in object detection for two reasons. First it is more difficult to design plausible augmentation strategies for object detection than for classification, because one must handle the complexity of bounding boxes if geometric transformations are applied. Secondly, data augmentation attracts less research attention perhaps because it is believed to add less value and to transfer poorly compared to advances in network architectures.This paper serves two main purposes. First, we propose to use AutoAugment [3] to design better data augmentation strategies for object detection because it can address the difficulty of designing them. Second, we use the method to assess the value of data augmentation in object detection and compare it against the value of architectures. Our investigation into data augmentation for object detection identifies two surprising results. First, by changing the data augmentation strategy to our method, AutoAugment for detection, we can improve RetinaNet with a ResNet-50 backbone from 36.7 to 39.0 mAP on COCO, a difference of +2.3 mAP. This gain exceeds the gain achieved by switching the backbone from ResNet-50 to ResNet-101 (+2.1 mAP), which incurs additional training and inference costs. The second surprising finding is that our strategies found on the COCO dataset transfer well to the PASCAL dataset to improve accuracy by +2.7 mAP. These results together with our systematic studies of data augmentation call into question previous assumptions about the role and transferability of architectures versus data augmentation. In particular, changing the augmentation may lead to performance gains that are equally transferable as changing the underlying architecture.

Efficient search is a core issue in Neural Architecture Search (NAS). It is difficult for conventional NAS algorithms to directly search the architectures on large-scale tasks like ImageNet. In general, the cost of GPU hours for NAS grows with regard to training dataset size and candidate set size. One common way is searching on a smaller proxy dataset (e.g., CIFAR-10) and then transferring to the target task (e.g., ImageNet). These architectures optimized on proxy data are not guaranteed to be optimal on the target task. Another common way is learning with a smaller candidate set, which may require expert knowledge and indeed betrays the essence of NAS. In this paper, we present DA-NAS that can directly search the architecture for large-scale target tasks while allowing a large candidate set in a more efficient manner. Our method is based on an interesting observation that the learning speed for blocks in deep neural networks is related to the difficulty of recognizing distinct categories. We carefully design a progressive data adapted pruning strategy for efficient architecture search. It will quickly trim low performed blocks on a subset of target dataset (e.g., easy classes), and then gradually find the best blocks on the whole target dataset. At this time, the original candidate set becomes as compact as possible, providing a faster search in the target task. Experiments on ImageNet verify the effectiveness of our approach. It is $$\mathbf{2} \varvec{\times }$$ 2 × faster than previous methods while the accuracy is currently state-of-the-art, at 76.2% under small FLOPs constraint. It supports an argument search space (i.e., more candidate blocks) to efficiently search the best-performing architecture.

Humans have a remarkable capacity to draw parallels between concepts, generalising their experience to new domains. This skill is essential to solving the visual problems featured in the RAVEN and PGM datasets, yet, previous papers have scarcely tested how well models generalise across tasks. Additionally, we encounter a critical issue that allows existing models to inadvertently ‘cheat’ problems in RAVEN. We therefore propose a simple workaround to resolve this issue, and focus the conversation on generalisation performance, as this was severely affected in the process. We revise the existing evaluation, and introduce two relational models, Rel-Base and Rel-AIR, that significantly improve this performance. To our knowledge, Rel-AIR is the first method to employ unsupervised scene decomposition in solving abstract visual reasoning problems, and along with Rel-Base, sets states-of-the-art for image-only reasoning and generalisation across both RAVEN and PGM.

Effectively keeping boundary of the mask complete is important in instance segmentation. In this task, many works segment instance based on a bounding box from the box head, which means the quality of the detection also affects the completeness of the mask. To circumvent this issue, we propose a fully convolutional box head and a supervised edge attention module in mask head. The box head contains one new IoU prediction branch. It learns association between object features and detected bounding boxes to provide more accurate bounding boxes for segmentation. The edge attention module utilizes attention mechanism to highlight object and suppress background noise, and a supervised branch is devised to guide the network to focus on the edge of instances precisely. To evaluate the effectiveness, we conduct experiments on COCO dataset. Without bells and whistles, our approach achieves impressive and robust improvement compared to baseline models. Code is at https://github.com//IPIU-detection/SEANet.

Domain adaptation (DA) has been a fundamental building block for Transfer Learning (TL) which assumes that source and target domain share the same label space. A more general and realistic setting is that the label space of target domain is a subset of the source domain, as termed by Partial domain adaptation (PDA). Previous methods typically match the whole source domain to target domain, which causes negative transfer due to the source-negative classes in source domain that does not exist in target domain. In this paper, a novel Discriminative Partial Domain Adversarial Network (DPDAN) is developed. We first propose to use hard binary weighting to differentiate the source-positive and source-negative samples in the source domain. The source-positive samples are those with labels shared by two domains, while the rest in the source domain are treated as source-negative samples. Based on the above binary relabeling strategy, our algorithm maximizes the distribution divergence between source-negative samples and all the others (source-positive and target samples), meanwhile minimizes domain shift between source-positive samples and target domain to obtain discriminative domain-invariant features. We empirically verify DPDAN can effectively reduce the negative transfer caused by source-negative classes, and also theoretically show it decreases negative transfer caused by domain shift. Experiments on four benchmark domain adaptation datasets show DPDAN consistently outperforms state-of-the-art methods.

Hyperspectral unmixing is an important remote sensing task with applications including material identification and analysis. Characteristic spectral features make many pure materials identifiable from their visible-to-infrared spectra, but quantifying their presence within a mixture is a challenging task due to nonlinearities and factors of variation. In this paper, spectral variation is considered from a physics-based approach and incorporated into an end-to-end spectral unmixing algorithm via differentiable programming. The dispersion model is introduced to simulate realistic spectral variation, and an efficient method to fit the parameters is presented. Then, this dispersion model is utilized as a generative model within an analysis-by-synthesis spectral unmixing algorithm. Further, a technique for inverse rendering using a convolutional neural network to predict parameters of the generative model is introduced to enhance performance and speed when training data is available. Results achieve state-of-the-art on both infrared and visible-to-near-infrared (VNIR) datasets, and show promise for the synergy between physics-based models and deep learning in hyperspectral unmixing in the future.

Although human face recognition has achieved exceptional success driven by deep learning, animal face recognition (AFR) is still a research field that received less attention. Due to the big challenge in collecting large-scale animal face datasets, it is difficult to train a high-precision AFR model from scratch. In this work, we propose a novel Residual InterSpecies Equivariant Network (RiseNet) to deal with the animal face recognition task with limited training samples. First, we formulate a module called residual inter-species feature equivariant to make the feature distribution of animals face closer to the human. Second, according to the structural characteristic of animal face, the features of the upper and lower half faces are learned separately. We present an animal facial feature fusion module to treat the features of the lower half face as additional information, which improves the proposed RiseNet performance. Besides, an animal face alignment strategy is designed for the preprocessing of the proposed network, which further aligns with the human face image. Extensive experiments on two benchmarks show that our method is effective and outperforms the state-of-the-arts.

Completing a corrupted image with correct structures and reasonable textures for a mixed scene remains an elusive challenge. Since the missing hole in a mixed scene of a corrupted image often contains various semantic information, conventional two-stage approaches utilizing structural information often lead to the problem of unreliable structural prediction and ambiguous image texture generation. In this paper, we propose a Semantic Guidance and Evaluation Network (SGE-Net) to iteratively update the structural priors and the inpainted image in an interplay framework of semantics extraction and image inpainting. It utilizes semantic segmentation map as guidance in each scale of inpainting, under which location-dependent inferences are re-evaluated, and, accordingly, poorly-inferred regions are refined in subsequent scales. Extensive experiments on real-world images of mixed scenes demonstrated the superiority of our proposed method over state-of-the-art approaches, in terms of clear boundaries and photo-realistic textures.

Learning associations across modalities is critical for robust multimodal reasoning, especially when a modality may be missing during inference. In this paper, we study this problem in the context of audio-conditioned visual synthesis – a task that is important, for example, in occlusion reasoning. Specifically, our goal is to generate future video frames and their motion dynamics conditioned on audio and a few past frames. To tackle this problem, we present Sound2Sight, a deep variational encoder-decoder framework, that is trained to learn a per frame stochastic prior conditioned on a joint embedding of audio and past frames. This embedding is learned via a multi-head attention-based audio-visual transformer encoder. The learned prior is then sampled to further condition a video forecasting module to generate future frames. The stochastic prior allows the model to sample multiple plausible futures that are consistent with the provided audio and the past context. Moreover, to improve the quality and coherence of the generated frames, we propose a multimodal discriminator that differentiates between a synthesized and a real audio-visual clip. We empirically evaluate our approach, vis-á-vis closely-related prior methods, on two new datasets viz. (i) Multimodal Stochastic Moving MNIST with a Surprise Obstacle, (ii) Youtube Paintings; as well as on the existing Audio-Set Drums dataset. Our extensive experiments demonstrate that Sound2Sight significantly outperforms the state of the art in the generated video quality, while also producing diverse video content.

In this paper, we propose a new deep architecture for fusing camera and LiDAR sensors for 3D object detection. Because the camera and LiDAR sensor signals have different characteristics and distributions, fusing these two modalities is expected to improve both the accuracy and robustness of 3D object detection. One of the challenges presented by the fusion of cameras and LiDAR is that the spatial feature maps obtained from each modality are represented by significantly different views in the camera and world coordinates; hence, it is not an easy task to combine two heterogeneous feature maps without loss of information. To address this problem, we propose a method called 3D-CVF that combines the camera and LiDAR features using the cross-view spatial feature fusion strategy. First, the method employs auto-calibrated projection, to transform the 2D camera features to a smooth spatial feature map with the highest correspondence to the LiDAR features in the bird’s eye view (BEV) domain. Then, a gated feature fusion network is applied to use the spatial attention maps to mix the camera and LiDAR features appropriately according to the region. Next, camera-LiDAR feature fusion is also achieved in the subsequent proposal refinement stage. The low-level LiDAR features and camera features are separately pooled using region of interest (RoI)-based feature pooling and fused with the joint camera-LiDAR features for enhanced proposal refinement. Our evaluation, conducted on the KITTI and nuScenes 3D object detection datasets, demonstrates that the camera-LiDAR fusion offers significant performance gain over the LiDAR-only baseline and that the proposed 3D-CVF achieves state-of-the-art performance in the KITTI benchmark.

Label noise is increasingly prevalent in datasets acquired from noisy channels. Existing approaches that detect and remove label noise generally rely on some form of supervision, which is not scalable and error-prone. In this paper, we propose NoiseRank, for unsupervised label noise reduction using Markov Random Fields (MRF). We construct a dependence model to estimate the posterior probability of an instance being incorrectly labeled given the dataset, and rank instances based on their estimated probabilities. Our method i) does not require supervision from ground-truth labels or priors on label or noise distribution, ii) is interpretable by design, enabling transparency in label noise removal, iii) is agnostic to classifier architecture/optimization framework and content modality. These advantages enable wide applicability in real noise settings, unlike prior works constrained by one or more conditions. NoiseRank improves state-of-the-art classification on Food101-N ( $$\sim $$ ∼ 20% noise), and is effective on high noise Clothing-1M ( $$\sim $$ ∼ 40% noise).

Conventional supervised super-resolution (SR) approaches are trained with massive external SR datasets but fail to exploit desirable properties of the given test image. On the other hand, self-supervised SR approaches utilize the internal information within a test image but suffer from computational complexity in run-time. In this work, we observe the opportunity for further improvement of the performance of single-image super-resolution (SISR) without changing the architecture of conventional SR networks by practically exploiting additional information given from the input image. In the training stage, we train the network via meta-learning; thus, the network can quickly adapt to any input image at test time. Then, in the test stage, parameters of this meta-learned network are rapidly fine-tuned with only a few iterations by only using the given low-resolution image. The adaptation at the test time takes full advantage of patch-recurrence property observed in natural images. Our method effectively handles unknown SR kernels and can be applied to any existing model. We demonstrate that the proposed model-agnostic approach consistently improves the performance of conventional SR networks on various benchmark SR datasets.

This paper proposes a novel deep convolutional model, Tri-Points Based Line Segment Detector (TP-LSD), to detect line segments in an image at real-time speed. The previous related methods typically use the two-step strategy, relying on either heuristic post-process or extra classifier. To realize one-step detection with a faster and more compact model, we introduce the tri-points representation, converting the line segment detection to the end-to-end prediction of a root-point and two endpoints for each line segment. TP-LSD has two branches: tri-points extraction branch and line segmentation branch. The former predicts the heat map of root-points and the two displacement maps of endpoints. The latter segments the pixels on straight lines out from background. Moreover, the line segmentation map is reused in the first branch as structural prior. We propose an additional novel evaluation metric and evaluate our method on Wireframe and YorkUrban datasets, demonstrating not only the competitive accuracy compared to the most recent methods, but also the real-time run speed up to 78 FPS with the $$320\times 320$$ 320 × 320 input.