Computer Vision – ECCV 2018

In this paper, we present a novel deep learning framework that derives discriminative local descriptors for 3D surface shapes. In contrast to previous convolutional neural networks (CNNs) that rely on rendering multi-view images or extracting intrinsic shape properties, we parameterize the multi-scale localized neighborhoods of a keypoint into regular 2D grids, which are termed as ‘geometry images’. The benefits of such geometry images include retaining sufficient geometric information, as well as allowing the usage of standard CNNs. Specifically, we leverage a triplet network to perform deep metric learning, which takes a set of triplets as input, and a newly designed triplet loss function is minimized to distinguish between similar and dissimilar pairs of keypoints. At the testing stage, given a geometry image of a point of interest, our network outputs a discriminative local descriptor for it. Experimental results for non-rigid shape matching on several benchmarks demonstrate the superior performance of our learned descriptors over traditional descriptors and the state-of-the-art learning-based alternatives.

Most of existing image denoising methods assume the corrupted noise to be additive white Gaussian noise (AWGN). However, the realistic noise in real-world noisy images is much more complex than AWGN, and is hard to be modeled by simple analytical distributions. As a result, many state-of-the-art denoising methods in literature become much less effective when applied to real-world noisy images captured by CCD or CMOS cameras. In this paper, we develop a trilateral weighted sparse coding (TWSC) scheme for robust real-world image denoising. Specifically, we introduce three weight matrices into the data and regularization terms of the sparse coding framework to characterize the statistics of realistic noise and image priors. TWSC can be reformulated as a linear equality-constrained problem and can be solved by the alternating direction method of multipliers. The existence and uniqueness of the solution and convergence of the proposed algorithm are analyzed. Extensive experiments demonstrate that the proposed TWSC scheme outperforms state-of-the-art denoising methods on removing realistic noise.

The automatic evaluation of image descriptions is an intricate task, and it is highly important in the development and fine-grained analysis of captioning systems. Existing metrics to automatically evaluate image captioning systems fail to achieve a satisfactory level of correlation with human judgements at the sentence level. Moreover, these metrics, unlike humans, tend to focus on specific aspects of quality, such as the n-gram overlap or the semantic meaning. In this paper, we present the first learning-based metric to evaluate image captions. Our proposed framework enables us to incorporate both lexical and semantic information into a single learned metric. This results in an evaluator that takes into account various linguistic features to assess the caption quality. The experiments we performed to assess the proposed metric, show improvements upon the state of the art in terms of correlation with human judgements and demonstrate its superior robustness to distractions.

Video object segmentation is challenging yet important in a wide variety of applications for video analysis. Recent works formulate video object segmentation as a prediction task using deep nets to achieve appealing state-of-the-art performance. Due to the formulation as a prediction task, most of these methods require fine-tuning during test time, such that the deep nets memorize the appearance of the objects of interest in the given video. However, fine-tuning is time-consuming and computationally expensive, hence the algorithms are far from real time. To address this issue, we develop a novel matching based algorithm for video object segmentation. In contrast to memorization based classification techniques, the proposed approach learns to match extracted features to a provided template without memorizing the appearance of the objects. We validate the effectiveness and the robustness of the proposed method on the challenging DAVIS-16, DAVIS-17, Youtube-Objects and JumpCut datasets. Extensive results show that our method achieves comparable performance without fine-tuning and is much more favorable in terms of computational time.

Current two-stage object detectors, which consists of a region proposal stage and a refinement stage, may produce unreliable results due to ill-localized proposed regions. To address this problem, we propose a context refinement algorithm that explores rich contextual information to better refine each proposed region. In particular, we first identify neighboring regions that may contain useful contexts and then perform refinement based on the extracted and unified contextual information. In practice, our method effectively improves the quality of the final detection results as well as region proposals. Empirical studies show that context refinement yields substantial and consistent improvements over different baseline detectors. Moreover, the proposed algorithm brings around 3% performance gain on PASCAL VOC benchmark and around 6% gain on MS COCO benchmark respectively.

Deep neural networks have enjoyed remarkable success for various vision tasks, however it remains challenging to apply CNNs to domains lacking a regular underlying structures such as 3D point clouds. Towards this we propose a novel convolutional architecture, termed SpiderCNN, to efficiently extract geometric features from point clouds. SpiderCNN is comprised of units called SpiderConv, which extend convolutional operations from regular grids to irregular point sets that can be embedded in $$\mathbb {R}^n$$ , by parametrizing a family of convolutional filters. We design the filter as a product of a simple step function that captures local geodesic information and a Taylor polynomial that ensures the expressiveness. SpiderCNN inherits the multi-scale hierarchical architecture from classical CNNs, which allows it to extract semantic deep features. Experiments on ModelNet40 demonstrate that SpiderCNN achieves state-of-the-art accuracy $$92.4 \%$$ on standard benchmarks, and shows competitive performance on segmentation task.

Diverse input data modalities can provide complementary cues for several tasks, usually leading to more robust algorithms and better performance. However, while a (training) dataset could be accurately designed to include a variety of sensory inputs, it is often the case that not all modalities are available in real life (testing) scenarios, where a model has to be deployed. This raises the challenge of how to learn robust representations leveraging multimodal data in the training stage, while considering limitations at test time, such as noisy or missing modalities. This paper presents a new approach for multimodal video action recognition, developed within the unified frameworks of distillation and privileged information, named generalized distillation. Particularly, we consider the case of learning representations from depth and RGB videos, while relying on RGB data only at test time. We propose a new approach to train an hallucination network that learns to distill depth features through multiplicative connections of spatiotemporal representations, leveraging soft labels and hard labels, as well as distance between feature maps. We report state-of-the-art results on video action classification on the largest multimodal dataset available for this task, the NTU RGB+D, as well as on the UWA3DII and Northwestern-UCLA.

Explanations of the decisions made by a deep neural network are important for human end-users to be able to understand and diagnose the trustworthiness of the system. Current neural networks used for visual recognition are generally used as black boxes that do not provide any human interpretable justification for a prediction. In this work we propose a new framework called Interpretable Basis Decomposition for providing visual explanations for classification networks. By decomposing the neural activations of the input image into semantically interpretable components pre-trained from a large concept corpus, the proposed framework is able to disentangle the evidence encoded in the activation feature vector, and quantify the contribution of each piece of evidence to the final prediction. We apply our framework for providing explanations to several popular networks for visual recognition, and show it is able to explain the predictions given by the networks in a human-interpretable way. The human interpretability of the visual explanations provided by our framework and other recent explanation methods is evaluated through Amazon Mechanical Turk, showing that our framework generates more faithful and interpretable explanations (The code and data are available at https://github.com/CSAILVision/IBD ).

Domain adversarial learning aligns the feature distributions across the source and target domains in a two-player minimax game. Existing domain adversarial networks generally assume identical label space across different domains. In the presence of big data, there is strong motivation of transferring deep models from existing big domains to unknown small domains. This paper introduces partial domain adaptation as a new domain adaptation scenario, which relaxes the fully shared label space assumption to that the source label space subsumes the target label space. Previous methods typically match the whole source domain to the target domain, which are vulnerable to negative transfer for the partial domain adaptation problem due to the large mismatch between label spaces. We present Partial Adversarial Domain Adaptation (PADA), which simultaneously alleviates negative transfer by down-weighing the data of outlier source classes for training both source classifier and domain adversary, and promotes positive transfer by matching the feature distributions in the shared label space. Experiments show that PADA exceeds state-of-the-art results for partial domain adaptation tasks on several datasets.

The large volume of video content and high viewing frequency demand automatic video summarization algorithms, of which a key property is the capability of modeling diversity. If videos are lengthy like hours-long egocentric videos, it is necessary to track the temporal structures of the videos and enforce local diversity. The local diversity refers to that the shots selected from a short time duration are diverse but visually similar shots are allowed to co-exist in the summary if they appear far apart in the video. In this paper, we propose a novel probabilistic model, built upon SeqDPP, to dynamically control the time span of a video segment upon which the local diversity is imposed. In particular, we enable SeqDPP to learn to automatically infer how local the local diversity is supposed to be from the input video. The resulting model is extremely involved to train by the hallmark maximum likelihood estimation (MLE), which further suffers from the exposure bias and non-differentiable evaluation metrics. To tackle these problems, we instead devise a reinforcement learning algorithm for training the proposed model. Extensive experiments verify the advantages of our model and the new learning algorithm over MLE-based methods.

Accurately localising object proposals is an important precondition for high detection rate for the state-of-the-art object detection frameworks. The accuracy of an object detection method has been shown highly related to the average recall (AR) of the proposals. In this work, we propose an advanced object proposal network in favour of translation-invariance for objectness classification, translation-variance for bounding box regression, large effective receptive fields for capturing global context and scale-invariance for dealing with a range of object sizes from extremely small to large. The design of the network architecture aims to be simple while being effective and with real-time performance. Without bells and whistles the proposed object proposal network significantly improves the AR at 1,000 proposals by $$35\%$$ and $$45\%$$ on PASCAL VOC and COCO dataset respectively and has a fast inference time of 44.8 ms for input image size of $$640^{2}$$ . Empirical studies have also shown that the proposed method is class-agnostic to be generalised for general object proposal.

Weight pruning methods for deep neural networks (DNNs) have been investigated recently, but prior work in this area is mainly heuristic, iterative pruning, thereby lacking guarantees on the weight reduction ratio and convergence time. To mitigate these limitations, we present a systematic weight pruning framework of DNNs using the alternating direction method of multipliers (ADMM). We first formulate the weight pruning problem of DNNs as a nonconvex optimization problem with combinatorial constraints specifying the sparsity requirements, and then adopt the ADMM framework for systematic weight pruning. By using ADMM, the original nonconvex optimization problem is decomposed into two subproblems that are solved iteratively. One of these subproblems can be solved using stochastic gradient descent, the other can be solved analytically. Besides, our method achieves a fast convergence rate.The weight pruning results are very promising and consistently outperform the prior work. On the LeNet-5 model for the MNIST data set, we achieve 71.2 $$\times $$ weight reduction without accuracy loss. On the AlexNet model for the ImageNet data set, we achieve 21 $$\times $$ weight reduction without accuracy loss. When we focus on the convolutional layer pruning for computation reductions, we can reduce the total computation by five times compared with the prior work (achieving a total of 13.4 $$\times $$ weight reduction in convolutional layers). Our models and codes are released at https://github.com/KaiqiZhang/admm-pruning .

In recent deep online and near-online multi-object tracking approaches, a difficulty has been to incorporate long-term appearance models to efficiently score object tracks under severe occlusion and multiple missing detections. In this paper, we propose a novel recurrent network model, the Bilinear LSTM, in order to improve the learning of long-term appearance models via a recurrent network. Based on intuitions drawn from recursive least squares, Bilinear LSTM stores building blocks of a linear predictor in its memory, which is then coupled with the input in a multiplicative manner, instead of the additive coupling in conventional LSTM approaches. Such coupling resembles an online learned classifier/regressor at each time step, which we have found to improve performances in using LSTM for appearance modeling. We also propose novel data augmentation approaches to efficiently train recurrent models that score object tracks on both appearance and motion. We train an LSTM that can score object tracks based on both appearance and motion and utilize it in a multiple hypothesis tracking framework. In experiments, we show that with our novel LSTM model, we achieved state-of-the-art performance on near-online multiple object tracking on the MOT 2016 and MOT 2017 benchmarks.

In this paper, we propose a novel method to compress CNNs by reconstructing the network from a small set of spatial convolution kernels. Starting from a pre-trained model, we extract representative 2D kernel centroids using k-means clustering. Each centroid replaces the corresponding kernels of the same cluster, and we use indexed representations instead of saving whole kernels. Kernels in the same cluster share their weights, and we fine-tune the model while keeping the compressed state. Furthermore, we also suggest an efficient way of removing redundant calculations in the compressed convolutional layers. We experimentally show that our technique works well without harming the accuracy of widely-used CNNs. Also, our ResNet-18 even outperforms its uncompressed counterpart at ILSVRC2012 classification task with over 10x compression ratio.

Fine-grained visual categorization (FGVC) is challenging due in part to the fact that it is often difficult to acquire an enough number of training samples. To employ large models for FGVC without suffering from overfitting, existing methods usually adopt a strategy of pre-training the models using a rich set of auxiliary data, followed by fine-tuning on the target FGVC task. However, the objective of pre-training does not take the target task into account, and consequently such obtained models are suboptimal for fine-tuning. To address this issue, we propose in this paper a new deep FGVC model termed MetaFGNet. Training of MetaFGNet is based on a novel regularized meta-learning objective, which aims to guide the learning of network parameters so that they are optimal for adapting to the target FGVC task. Based on MetaFGNet, we also propose a simple yet effective scheme for selecting more useful samples from the auxiliary data. Experiments on benchmark FGVC datasets show the efficacy of our proposed method.

The requirement of large amounts of annotated images has become one grand challenge while training deep neural network models for various visual detection and recognition tasks. This paper presents a novel image synthesis technique that aims to generate a large amount of annotated scene text images for training accurate and robust scene text detection and recognition models. The proposed technique consists of three innovative designs. First, it realizes “semantic coherent” synthesis by embedding texts at semantically sensible regions within the background image, where the semantic coherence is achieved by leveraging the semantic annotations of objects and image regions that have been created in the prior semantic segmentation research. Second, it exploits visual saliency to determine the embedding locations within each semantic sensible region, which coincides with the fact that texts are often placed around homogeneous regions for better visibility in scenes. Third, it designs an adaptive text appearance model that determines the color and brightness of embedded texts by learning from the feature of real scene text images adaptively. The proposed technique has been evaluated over five public datasets and the experiments show its superior performance in training accurate and robust scene text detection and recognition models.

In this paper, we propose a simple and general framework for training very tiny CNNs (e.g. VGG with the number of channels reduced to $$\frac{1}{32}$$ ) for object detection. Due to limited representation ability, it is challenging to train very tiny networks for complicated tasks like detection. To the best of our knowledge, our method, called Quantization Mimic, is the first one focusing on very tiny networks. We utilize two types of acceleration methods: mimic and quantization. Mimic improves the performance of a student network by transfering knowledge from a teacher network. Quantization converts a full-precision network to a quantized one without large degradation of performance. If the teacher network is quantized, the search scope of the student network will be smaller. Using this feature of the quantization, we propose Quantization Mimic. It first quantizes the large network, then mimic a quantized small network. The quantization operation can help student network to better match the feature maps from teacher network. To evaluate our approach, we carry out experiments on various popular CNNs including VGG and Resnet, as well as different detection frameworks including Faster R-CNN and R-FCN. Experiments on Pascal VOC and WIDER FACE verify that our Quantization Mimic algorithm can be applied on various settings and outperforms state-of-the-art model acceleration methods given limited computing resouces.

Sum-of-squares objective functions are very popular in computer vision algorithms. However, these objective functions are not always easy to optimize. The underlying assumptions made by solvers are often not satisfied and many problems are inherently ill-posed. In this paper, we propose a neural nonlinear least squares optimization algorithm which learns to effectively optimize these cost functions even in the presence of adversities. Unlike traditional approaches, the proposed solver requires no hand-crafted regularizers or priors as these are implicitly learned from the data. We apply our method to the problem of motion stereo ie. jointly estimating the motion and scene geometry from pairs of images of a monocular sequence. We show that our learned optimizer is able to efficiently and effectively solve this challenging optimization problem.

In this paper we propose a novel decomposition method based on filter group approximation, which can significantly reduce the redundancy of deep convolutional neural networks (CNNs) while maintaining the majority of feature representation. Unlike other low-rank decomposition algorithms which operate on spatial or channel dimension of filters, our proposed method mainly focuses on exploiting the filter group structure for each layer. For several commonly used CNN models, including VGG and ResNet, our method can reduce over 80% floating-point operations (FLOPs) with less accuracy drop than state-of-the-art methods on various image classification datasets. Besides, experiments demonstrate that our method is conducive to alleviating degeneracy of the compressed network, which hurts the convergence and performance of the network.

This paper proposes a real-time dynamic scene reconstruction method capable of reproducing the motion, geometry, and segmentation simultaneously given live depth stream from a single RGB-D camera. Our approach fuses geometry frame by frame and uses a segmentation-enhanced node graph structure to drive the deformation of geometry in registration step. A two-level node motion optimization is proposed. The optimization space of node motions and the range of physically-plausible deformations are largely reduced by taking advantage of the articulated motion prior, which is solved by an efficient node graph segmentation method. Compared to previous fusion-based dynamic scene reconstruction methods, our experiments show robust and improved reconstruction results for tangential and occluded motions.

Solving a multi-labeling problem with a convex penalty can be achieved in polynomial time if the label set is totally ordered. In this paper we propose a generalization to partially ordered sets. To this end, we assume that the label set is the Cartesian product of totally ordered sets and the convex prior is separable. For this setting we introduce a general combinatorial optimization framework that provides an approximate solution. More specifically, we first construct a graph whose minimal cut provides a lower bound to our energy. The result of this relaxation is then used to get a feasible solution via classical move-making cuts. To speed up the optimization, we propose an efficient coarse-to-fine approach over the label space. We demonstrate the proposed framework through extensive experiments for optical flow estimation.

We propose a novel attention mechanism to enhance Convolutional Neural Networks for fine-grained recognition. It learns to attend to lower-level feature activations without requiring part annotations and uses these activations to update and rectify the output likelihood distribution. In contrast to other approaches, the proposed mechanism is modular, architecture-independent and efficient both in terms of parameters and computation required. Experiments show that networks augmented with our approach systematically improve their classification accuracy and become more robust to clutter. As a result, Wide Residual Networks augmented with our proposal surpasses the state of the art classification accuracies in CIFAR-10, the Adience gender recognition task, Stanford dogs, and UEC Food-100.

Although weight and activation quantization is an effective approach for Deep Neural Network (DNN) compression and has a lot of potentials to increase inference speed leveraging bit-operations, there is still a noticeable gap in terms of prediction accuracy between the quantized model and the full-precision model. To address this gap, we propose to jointly train a quantized, bit-operation-compatible DNN and its associated quantizers, as opposed to using fixed, handcrafted quantization schemes such as uniform or logarithmic quantization. Our method for learning the quantizers applies to both network weights and activations with arbitrary-bit precision, and our quantizers are easy to train. The comprehensive experiments on CIFAR-10 and ImageNet datasets show that our method works consistently well for various network structures such as AlexNet, VGG-Net, GoogLeNet, ResNet, and DenseNet, surpassing previous quantization methods in terms of accuracy by an appreciable margin. Code available at https://github.com/Microsoft/LQ-Nets .

Supervised learning techniques have shown substantial progress on video summarization. State-of-the-art approaches mostly regard the predicted summary and the human summary as two sequences (sets), and minimize discriminative losses that measure element-wise discrepancy. Such training objectives do not explicitly model how well the predicted summary preserves semantic information in the video. Moreover, those methods often demand a large amount of human generated summaries. In this paper, we propose a novel sequence-to-sequence learning model to address these deficiencies. The key idea is to complement the discriminative losses with another loss which measures if the predicted summary preserves the same information as in the original video. To this end, we propose to augment standard sequence learning models with an additional “retrospective encoder” that embeds the predicted summary into an abstract semantic space. The embedding is then compared to the embedding of the original video in the same space. The intuition is that both embeddings ought to be close to each other for a video and its corresponding summary. Thus our approach adds to the discriminative loss a metric learning loss that minimizes the distance between such pairs while maximizing the distances between unmatched ones. One important advantage is that the metric learning loss readily allows learning from videos without human generated summaries. Extensive experimental results show that our model outperforms existing ones by a large margin in both supervised and semi-supervised settings.

Deep neural network compression has the potential to bring modern resource-hungry deep networks to resource-limited devices. However, in many of the most compelling deployment scenarios of compressed deep networks, the operational constraints matter: for example, a pedestrian detection network on a self-driving car may have to satisfy a latency constraint for safe operation. We propose the first principled treatment of deep network compression under operational constraints. We formulate the compression learning problem from the perspective of constrained Bayesian optimization, and introduce a cooling (annealing) strategy to guide the network compression towards the target constraints. Experiments on ImageNet demonstrate the value of modelling constraints directly in network compression.

An ever increasing amount of our digital communication, media consumption, and content creation revolves around videos. We share, watch, and archive many aspects of our lives through them, all of which are powered by strong video compression. Traditional video compression is laboriously hand designed and hand optimized. This paper presents an alternative in an end-to-end deep learning codec. Our codec builds on one simple idea: Video compression is repeated image interpolation. It thus benefits from recent advances in deep image interpolation and generation. Our deep video codec outperforms today’s prevailing codecs, such as H.261, MPEG-4 Part 2, and performs on par with H.264.

Human motion prediction, forecasting human motion in a few milliseconds conditioning on a historical 3D skeleton sequence, is a long-standing problem in computer vision and robotic vision. Existing forecasting algorithms rely on extensive annotated motion capture data and are brittle to novel actions. This paper addresses the problem of few-shot human motion prediction, in the spirit of the recent progress on few-shot learning and meta-learning. More precisely, our approach is based on the insight that having a good generalization from few examples relies on both a generic initial model and an effective strategy for adapting this model to novel tasks. To accomplish this, we propose proactive and adaptive meta-learning (PAML) that introduces a novel combination of model-agnostic meta-learning and model regression networks and unifies them into an integrated, end-to-end framework. By doing so, our meta-learner produces a generic initial model through aggregating contextual information from a variety of prediction tasks, while effectively adapting this model for use as a task-specific one by leveraging learning-to-learn knowledge about how to transform few-shot model parameters to many-shot model parameters. The resulting PAML predictor model significantly improves the prediction performance on the heavily benchmarked H3.6M dataset in the small-sample size regime.

Question answering is an important task for autonomous agents and virtual assistants alike and was shown to support the disabled in efficiently navigating an overwhelming environment. Many existing methods focus on observation-based questions, ignoring our ability to seamlessly combine observed content with general knowledge. To understand interactions with a knowledge base, a dataset has been introduced recently and keyword matching techniques were shown to yield compelling results despite being vulnerable to misconceptions due to synonyms and homographs. To address this issue, we develop a learning-based approach which goes straight to the facts via a learned embedding space. We demonstrate state-of-the-art results on the challenging recently introduced fact-based visual question answering dataset, outperforming competing methods by more than $$5\%$$ .

Despite the fact that nonlinear subspace learning techniques (e.g. manifold learning) have successfully applied to data representation, there is still room for improvement in explainability (explicit mapping), generalization (out-of-samples), and cost-effectiveness (linearization). To this end, a novel linearized subspace learning technique is developed in a joint and progressive way, called joint and progressive learning strategy (J-Play), with its application to multi-label classification. The J-Play learns high-level and semantically meaningful feature representation from high-dimensional data by (1) jointly performing multiple subspace learning and classification to find a latent subspace where samples are expected to be better classified; (2) progressively learning multi-coupled projections to linearly approach the optimal mapping bridging the original space with the most discriminative subspace; (3) locally embedding manifold structure in each learnable latent subspace. Extensive experiments are performed to demonstrate the superiority and effectiveness of the proposed method in comparison with previous state-of-the-art methods.

We introduce Spatial-Temporal Memory Networks for video object detection. At its core, a novel Spatial-Temporal Memory module (STMM) serves as the recurrent computation unit to model long-term temporal appearance and motion dynamics. The STMM’s design enables full integration of pretrained backbone CNN weights, which we find to be critical for accurate detection. Furthermore, in order to tackle object motion in videos, we propose a novel MatchTrans module to align the spatial-temporal memory from frame to frame. Our method produces state-of-the-art results on the benchmark ImageNet VID dataset, and our ablative studies clearly demonstrate the contribution of our different design choices. We release our code and models at http://fanyix.cs.ucdavis.edu/project/stmn/project.html .

The quick detection of specific substances in objects such as produce items via non-destructive visual cues is vital to ensuring the quality and safety of consumer products. At the same time, it is well-known that the fluorescence excitation-emission characteristics of many organic objects can serve as a kind of “fingerprint” for detecting the presence of specific substances in classification tasks such as determining if something is safe to consume. However, conventional capture of the fluorescence excitation-emission matrix can take on the order of minutes and can only be done for point measurements. In this paper, we propose a coded illumination approach whereby light spectra are learned such that key visual fluorescent features can be easily seen for material classification. We show that under a single coded illuminant, we can capture one RGB image and perform pixel-level classifications of materials at high accuracy. This is demonstrated through effective classification of different types of honey and alcohol using real images.

Recent studies have shown that deep neural networks can significantly improve the quality of single-image super-resolution. Current researches tend to use deeper convolutional neural networks to enhance performance. However, blindly increasing the depth of the network cannot ameliorate the network effectively. Worse still, with the depth of the network increases, more problems occurred in the training process and more training tricks are needed. In this paper, we propose a novel multi-scale residual network (MSRN) to fully exploit the image features, which outperform most of the state-of-the-art methods. Based on the residual block, we introduce convolution kernels of different sizes to adaptively detect the image features in different scales. Meanwhile, we let these features interact with each other to get the most efficacious image information, we call this structure Multi-scale Residual Block (MSRB). Furthermore, the outputs of each MSRB are used as the hierarchical features for global feature fusion. Finally, all these features are sent to the reconstruction module for recovering the high-quality image.

Autonomous vehicles require knowledge of the surrounding road layout, which can be predicted by state-of-the-art CNNs. This work addresses the current lack of data for determining lane instances, which are needed for various driving manoeuvres. The main issue is the time-consuming manual labelling process, typically applied per image. We notice that driving the car is itself a form of annotation. Therefore, we propose a semi-automated method that allows for efficient labelling of image sequences by utilising an estimated road plane in 3D based on where the car has driven and projecting labels from this plane into all images of the sequence. The average labelling time per image is reduced to 5 s and only an inexpensive dash-cam is required for data capture. We are releasing a dataset of 24,000 images and additionally show experimental semantic segmentation and instance segmentation results.

As deep learning methods form a critical part in commercially important applications such as autonomous driving and medical diagnostics, it is important to reliably detect out-of-distribution (OOD) inputs while employing these algorithms. In this work, we propose an OOD detection algorithm which comprises of an ensemble of classifiers. We train each classifier in a self-supervised manner by leaving out a random subset of training data as OOD data and the rest as in-distribution (ID) data. We propose a novel margin-based loss over the softmax output which seeks to maintain at least a margin m between the average entropy of the OOD and in-distribution samples. In conjunction with the standard cross-entropy loss, we minimize the novel loss to train an ensemble of classifiers. We also propose a novel method to combine the outputs of the ensemble of classifiers to obtain OOD detection score and class prediction. Overall, our method convincingly outperforms Hendrycks et al. [7] and the current state-of-the-art ODIN [13] on several OOD detection benchmarks.

Many recent energy-based methods for optical flow estimation rely on a good initialization that is typically provided by some kind of feature matching. So far, however, these initial matching approaches are rather general: They do not incorporate any additional information that could help to improve the accuracy or the robustness of the estimation. In particular, they do not exploit potential cues on the camera poses and the thereby induced rigid motion of the scene. In the present paper, we tackle this problem. To this end, we propose a novel structure-from-motion-aware PatchMatch approach that, in contrast to existing matching techniques, combines two hierarchical feature matching methods: a recent two-frame PatchMatch approach for optical flow estimation (general motion) and a specifically tailored three-frame PatchMatch approach for rigid scene reconstruction (SfM). While the motion PatchMatch serves as baseline with good accuracy, the SfM counterpart takes over at occlusions and other regions with insufficient information. Experiments with our novel SfM-aware PatchMatch approach demonstrate its usefulness. They not only show excellent results for all major benchmarks (KITTI 2012/2015, MPI Sintel), but also improvements up to 50% compared to a PatchMatch approach without structure information.

In this work we aim to develop a universal sketch grouper. That is, a grouper that can be applied to sketches of any category in any domain to group constituent strokes/segments into semantically meaningful object parts. The first obstacle to this goal is the lack of large-scale datasets with grouping annotation. To overcome this, we contribute the largest sketch perceptual grouping (SPG) dataset to date, consisting of 20, 000 unique sketches evenly distributed over 25 object categories. Furthermore, we propose a novel deep universal perceptual grouping model. The model is learned with both generative and discriminative losses. The generative losses improve the generalisation ability of the model to unseen object categories and datasets. The discriminative losses include a local grouping loss and a novel global grouping loss to enforce global grouping consistency. We show that the proposed model significantly outperforms the state-of-the-art groupers. Further, we show that our grouper is useful for a number of sketch analysis tasks including sketch synthesis and fine-grained sketch-based image retrieval (FG-SBIR).

Imagining a scene described in natural language with realistic layout and appearance of entities is the ultimate test of spatial, visual, and semantic world knowledge. Towards this goal, we present the Composition, Retrieval and Fusion Network (Craft), a model capable of learning this knowledge from video-caption data and applying it while generating videos from novel captions. Craft explicitly predicts a temporal-layout of mentioned entities (characters and objects), retrieves spatio-temporal entity segments from a video database and fuses them to generate scene videos. Our contributions include sequential training of components of Craft while jointly modeling layout and appearances, and losses that encourage learning compositional representations for retrieval. We evaluate Craft on semantic fidelity to caption, composition consistency, and visual quality. Craft outperforms direct pixel generation approaches and generalizes well to unseen captions and to unseen video databases with no text annotations. We demonstrate Craft on Flintstones (Flintstones is available at https://prior.allenai.org/projects/craft ), a new richly annotated video-caption dataset with over 25000 videos. For a glimpse of videos generated by Craft, see https://youtu.be/688Vv86n0z8 .

Urban zoning enables various applications in land use analysis and urban planning. As cities evolve, it is important to constantly update the zoning maps of cities to reflect urban pattern changes. This paper proposes a method for automatic urban zoning using higher-order Markov random fields (HO-MRF) built on multi-view imagery data including street-view photos and top-view satellite images. In the proposed HO-MRF, top-view satellite data is segmented via a multi-scale deep convolutional neural network (MS-CNN) and used in lower-order potentials. Street-view data with geo-tagged information is augmented in higher-order potentials. Various feature types for classifying street-view images were also investigated in our work. We evaluated the proposed method on a number of famous metropolises and provided in-depth analysis on technical issues.

Neural networks in the real domain have been studied for a long time and achieved promising results in many vision tasks for recent years. However, the extensions of the neural network models in other number fields and their potential applications are not fully-investigated yet. Focusing on color images, which can be naturally represented as quaternion matrices, we propose a quaternion convolutional neural network (QCNN) model to obtain more representative features. In particular, we re-design the basic modules like convolution layer and fully-connected layer in the quaternion domain, which can be used to establish fully-quaternion convolutional neural networks. Moreover, these modules are compatible with almost all deep learning techniques and can be plugged into traditional CNNs easily. We test our QCNN models in both color image classification and denoising tasks. Experimental results show that they outperform the real-valued CNNs with same structures.

Stereo relative pose problem lies at the core of stereo visual odometry systems that are used in many applications. In this work we present two minimal solvers for the stereo relative pose. We specifically consider the case when a minimal set consists of three point or line features and each of them has three known projections on two stereo cameras. We validate the importance of this formulation for practical purposes in our experiments with motion estimation. We then present a complete classification of minimal cases with three point or line correspondences each having three projections, and present two new solvers that can handle all such cases. We demonstrate a considerable effect from the integration of the new solvers into a visual SLAM system.

This paper presents novel techniques for recovering 3D dense scene flow, based on differential analysis of 4D light fields. The key enabling result is a per-ray linear equation, called the ray flow equation, that relates 3D scene flow to 4D light field gradients. The ray flow equation is invariant to 3D scene structure and applicable to a general class of scenes, but is underconstrained (3 unknowns per equation). Thus, additional constraints must be imposed to recover motion. We develop two families of scene flow algorithms by leveraging the structural similarity between ray flow and optical flow equations: local ‘Lucas-Kanade’ ray flow and global ‘Horn-Schunck’ ray flow, inspired by corresponding optical flow methods. We also develop a combined local-global method by utilizing the correspondence structure in the light fields. We demonstrate high precision 3D scene flow recovery for a wide range of scenarios, including rotation and non-rigid motion. We analyze the theoretical and practical performance limits of the proposed techniques via the light field structure tensor, a $$3 \times 3$$ 3 × 3 matrix that encodes the local structure of light fields. We envision that the proposed analysis and algorithms will lead to design of future light-field cameras that are optimized for motion sensing, in addition to depth sensing.

Neglecting the effects of rolling-shutter cameras for visual odometry (VO) severely degrades accuracy and robustness. In this paper, we propose a novel direct monocular VO method that incorporates a rolling-shutter model. Our approach extends direct sparse odometry which performs direct bundle adjustment of a set of recent keyframe poses and the depths of a sparse set of image points. We estimate the velocity at each keyframe and impose a constant-velocity prior for the optimization. In this way, we obtain a near real-time, accurate direct VO method. Our approach achieves improved results on challenging rolling-shutter sequences over state-of-the-art global-shutter VO.

Recently, style transfer has received a lot of attention. While much of this research has aimed at speeding up processing, the approaches are still lacking from a principled, art historical standpoint: a style is more than just a single image or an artist, but previous work is limited to only a single instance of a style or shows no benefit from more images. Moreover, previous work has relied on a direct comparison of art in the domain of RGB images or on CNNs pre-trained on ImageNet, which requires millions of labeled object bounding boxes and can introduce an extra bias, since it has been assembled without artistic consideration. To circumvent these issues, we propose a style-aware content loss, which is trained jointly with a deep encoder-decoder network for real-time, high-resolution stylization of images and videos. We propose a quantitative measure for evaluating the quality of a stylized image and also have art historians rank patches from our approach against those from previous work. These and our qualitative results ranging from small image patches to megapixel stylistic images and videos show that our approach better captures the subtle nature in which a style affects content.

We propose a novel direct visual odometry algorithm for micro-lens-array-based light field cameras. The algorithm calculates a detailed, semi-dense 3D point cloud of its environment. This is achieved by establishing probabilistic depth hypotheses based on stereo observations between the micro images of different recordings. Tracking is performed in a coarse-to-fine process, working directly on the recorded raw images. The tracking accounts for changing lighting conditions and utilizes a linear motion model to be more robust. A novel scale optimization framework is proposed. It estimates the scene scale, on the basis of keyframes, and optimizes the scale of the entire trajectory by filtering over multiple estimates. The method is tested based on a versatile dataset consisting of challenging indoor and outdoor sequences and is compared to state-of-the-art monocular and stereo approaches. The algorithm shows the ability to recover the absolute scale of the scene and significantly outperforms state-of-the-art monocular algorithms with respect to scale drifts.

We propose a neural approach for fusing an arbitrary-length burst of photographs suffering from severe camera shake and noise into a sharp and noise-free image. Our novel convolutional architecture has a simultaneous view of all frames in the burst, and by construction treats them in an order-independent manner. This enables it to effectively detect and leverage subtle cues scattered across different frames, while ensuring that each frame gets a full and equal consideration regardless of its position in the sequence. We train the network with richly varied synthetic data consisting of camera shake, realistic noise, and other common imaging defects. The method demonstrates consistent state of the art burst image restoration performance for highly degraded sequences of real-world images, and extracts accurate detail that is not discernible from any of the individual frames in isolation.

We introduce a novel RGB-D patch descriptor designed for detecting coplanar surfaces in SLAM reconstruction. The core of our method is a deep convolutional neural network that takes in RGB, depth, and normal information of a planar patch in an image and outputs a descriptor that can be used to find coplanar patches from other images. We train the network on 10 million triplets of coplanar and non-coplanar patches, and evaluate on a new coplanarity benchmark created from commodity RGB-D scans. Experiments show that our learned descriptor outperforms alternatives extended for this new task by a significant margin. In addition, we demonstrate the benefits of coplanarity matching in a robust RGBD reconstruction formulation. We find that coplanarity constraints detected with our method are sufficient to get reconstruction results comparable to state-of-the-art frameworks on most scenes, but outperform other methods on established benchmarks when combined with traditional keypoint matching.

We present an end-to-end deep learning architecture for depth map inference from multi-view images. In the network, we first extract deep visual image features, and then build the 3D cost volume upon the reference camera frustum via the differentiable homography warping. Next, we apply 3D convolutions to regularize and regress the initial depth map, which is then refined with the reference image to generate the final output. Our framework flexibly adapts arbitrary N-view inputs using a variance-based cost metric that maps multiple features into one cost feature. The proposed MVSNet is demonstrated on the large-scale indoor DTU dataset. With simple post-processing, our method not only significantly outperforms previous state-of-the-arts, but also is several times faster in runtime. We also evaluate MVSNet on the complex outdoor Tanks and Temples dataset, where our method ranks first before April 18, 2018 without any fine-tuning, showing the strong generalization ability of MVSNet.

In this paper we present ActiveStereoNet, the first deep learning solution for active stereo systems. Due to the lack of ground truth, our method is fully self-supervised, yet it produces precise depth with a subpixel precision of 1 / 30th of a pixel; it does not suffer from the common over-smoothing issues; it preserves the edges; and it explicitly handles occlusions. We introduce a novel reconstruction loss that is more robust to noise and texture-less patches, and is invariant to illumination changes. The proposed loss is optimized using a window-based cost aggregation with an adaptive support weight scheme. This cost aggregation is edge-preserving and smooths the loss function, which is key to allow the network to reach compelling results. Finally we show how the task of predicting invalid regions, such as occlusions, can be trained end-to-end without ground-truth. This component is crucial to reduce blur and particularly improves predictions along depth discontinuities. Extensive quantitatively and qualitatively evaluations on real and synthetic data demonstrate state of the art results in many challenging scenes.

In this paper, we present a framework for reconstructing a point-based 3D model of an object from a single-view image. We found distance metrics, like Chamfer distance, were used in previous work to measure the difference of two point sets and serve as the loss function in point-based reconstruction. However, such point-point loss does not constrain the 3D model from a global perspective. We propose adding geometric adversarial loss (GAL). It is composed of two terms where the geometric loss ensures consistent shape of reconstructed 3D models close to ground-truth from different viewpoints, and the conditional adversarial loss generates a semantically-meaningful point cloud. GAL benefits predicting the obscured part of objects and maintaining geometric structure of the predicted 3D model. Both the qualitative results and quantitative analysis manifest the generality and suitability of our method.

Monocular visual odometry approaches that purely rely on geometric cues are prone to scale drift and require sufficient motion parallax in successive frames for motion estimation and 3D reconstruction. In this paper, we propose to leverage deep monocular depth prediction to overcome limitations of geometry-based monocular visual odometry. To this end, we incorporate deep depth predictions into Direct Sparse Odometry (DSO) as direct virtual stereo measurements. For depth prediction, we design a novel deep network that refines predicted depth from a single image in a two-stage process. We train our network in a semi-supervised way on photoconsistency in stereo images and on consistency with accurate sparse depth reconstructions from Stereo DSO. Our deep predictions excel state-of-the-art approaches for monocular depth on the KITTI benchmark. Moreover, our Deep Virtual Stereo Odometry clearly exceeds previous monocular and deep-learning based methods in accuracy. It even achieves comparable performance to the state-of-the-art stereo methods, while only relying on a single camera.