CVPR 2018 Papers

We propose a network for Congested Scene Recognition called CSRNet to provide a data-driven and deep learning method that can understand highly congested scenes and perform accurate count estimation as well as present high-quality density maps. The proposed CSRNet is composed of two major components: a convolutional neural network (CNN) as the front-end for 2D feature extraction and a dilated CNN for the back-end, which uses dilated kernels to deliver larger reception fields and to replace pooling operations. CSRNet is an easy-trained model because of its pure convolutional structure. We demonstrate CSRNet on four datasets (ShanghaiTech dataset, the UCF_CC_50 dataset, the WorldEXPO'10 dataset, and the UCSD dataset) and we deliver the state-of-the-art performance. In the ShanghaiTech Part_B dataset, CSRNet achieves 47.3% lower Mean Absolute Error (MAE) than the previous state-of-the-art method. We extend the targeted applications for counting other objects, such as the vehicle in TRANCOS dataset. Results show that CSRNet significantly improves the output quality with 15.4% lower MAE than the previous state-of-the-art approach.

We present a neural architecture that takes as input a 2D or 3D shape and outputs a program that generates the shape. The instructions in our program are based on constructive solid geometry principles, i.e., a set of boolean operations on shape primitives defined recursively. Bottom-up techniques for this shape parsing task rely on primitive detection and are inherently slow since the search space over possible primitive combinations is large. In contrast, our model uses a recurrent neural network that parses the input shape in a top-down manner, which is significantly faster and yields a compact and easy-to-interpret sequence of modeling instructions. Our model is also more effective as a shape detector compared to existing state-of-the-art detection techniques. We finally demonstrate that our network can be trained on novel datasets without ground-truth program annotations through policy gradient techniques.

Removing the undesired reflections from images taken through the glass is of broad application to various computer vision tasks. Non-learning based methods utilize different handcrafted priors such as the separable sparse gradients caused by different levels of blurs, which often fail due to their limited description capability to the properties of real-world reflections. In this paper, we propose the Concurrent Reflection Removal Network (CRRN) to tackle this problem in a unified framework. Our network integrates image appearance information and multi-scale gradient information with human perception inspired loss function, and is trained on a new dataset with 3250 reflection images taken under diverse real-world scenes. Extensive experiments on a public benchmark dataset show that the proposed method performs favorably against state-of-the-art methods.

Semantic classes can be either things (objects with a well-defined shape, e.g. car, person) or stuff (amorphous background regions, e.g. grass, sky). While lots of classification and detection works focus on thing classes, less attention has been given to stuff classes. Nonetheless, stuff classes are important as they allow to explain important aspects of an image, including (1) scene type; (2) which thing classes are likely to be present and their location (through contextual reasoning); (3) physical attributes, material types and geometric properties of the scene. To understand stuff and things in context we introduce COCO-Stuff, which augments all 164K images of the COCO 2017 dataset with pixel-wise annotations for 91 stuff classes. We introduce an efficient stuff annotation protocol based on superpixels, which leverages the original thing annotations. We quantify the speed versus quality trade-off of our protocol and explore the relation between annotation time and boundary complexity. Furthermore, we use COCO-Stuff to analyze: (a) the importance of stuff and thing classes in terms of their surface cover and how frequently they are mentioned in image captions; (b) the spatial relations between stuff and things, highlighting the rich contextual relations that make our dataset unique; (c) the performance of a modern semantic segmentation method on stuff and thing classes, and whether stuff is easier to segment than things.

This paper addresses the problem of video object segmentation, where the initial object mask is given in the first frame of an input video. We propose a novel spatio-temporal Markov Random Field (MRF) model defined over pixels to handle this problem. Unlike conventional MRF models, the spatial dependencies among pixels in our model are encoded by a Convolutional Neural Network (CNN). Specifically, for a given object, the probability of a labeling to a set of spatially neighboring pixels can be predicted by a CNN trained for this specific object. As a result, higher-order, richer dependencies among pixels in the set can be implicitly modeled by the CNN. With temporal dependencies established by optical flow, the resulting MRF model combines both spatial and temporal cues for tackling video object segmentation. However, performing inference in the MRF model is very difficult due to the very high-order dependencies. To this end, we propose a novel CNN-embedded algorithm to perform approximate inference in the MRF. This algorithm proceeds by alternating between a temporal fusion step and a feed-forward CNN step. When initialized with an appearance-based one-shot segmentation CNN, our model outperforms the winning entries of the DAVIS 2017 Challenge, without resorting to model ensembling or any dedicated detectors.

Traditional single-grid and pyramidal B-spline parameterizations used in deformable image registration require users to specify control point spacing configurations capable of accurately capturing both global and complex local deformations. In many cases, such grid configurations are non-obvious and largely selected based on user experience. Recent regularization methods imposing sparsity upon the B-spline coefficients throughout simultaneous multi-grid optimization, however, have provided a promising means of determining suitable configurations automatically. Unfortunately, imposing sparsity on over-parameterized B-spline models is computationally expensive and introduces additional difficulties such as undesirable local minima in the B-spline coefficient optimization process. To overcome these difficulties in determining B-spline grid configurations, this paper investigates the use of convolutional neural networks (CNNs) to learn and infer expressive sparse multi-grid configurations prior to B-spline coefficient optimization. Experimental results show that multi-grid configurations produced in this fashion using our CNN based approach provide registration quality comparable to L1-norm constrained over-parameterizations in terms of exactness, while exhibiting significantly reduced computational requirements.

Most of the traditional work on intrinsic image decomposition rely on deriving priors about scene characteristics. On the other hand, recent research use deep learning models as in-and-out black box and do not consider the well-established, traditional image formation process as the basis of their intrinsic learning process. As a consequence, although current deep learning approaches show superior performance when considering quantitative benchmark results, traditional approaches are still dominant in achieving high qualitative results. In this paper, the aim is to exploit the best of the two worlds. A method is proposed that (1) is empowered by deep learning capabilities, (2) considers a physics-based reflection model to steer the learning process, and (3) exploits the traditional approach to obtain intrinsic images by exploiting reflectance and shading gradient information. The proposed model is fast to compute and allows for the integration of all intrinsic components. To train the new model, an object centered large-scale datasets with intrinsic ground-truth images are created. The evaluation results demonstrate that the new model outperforms existing methods. Visual inspection shows that the image formation loss function augments color reproduction and the use of gradient information produces sharper edges. Datasets, models and higher resolution images are available at https://ivi.fnwi.uva.nl/cv/retinet.

Deep neural networks enable state-of-the-art accuracy on visual recognition tasks such as image classification and object detection. However, modern deep networks contain millions of learned weights; a more efficient utilization of computation resources would assist in a variety of deployment scenarios, from embedded platforms with resource constraints to computing clusters running ensembles of networks. In this paper, we combine network pruning and weight quantization in a single learning framework that performs pruning and quantization jointly, and in parallel with fine-tuning. This allows us to take advantage of the complementary nature of pruning and quantization and to recover from premature pruning errors, which is not possible with current two-stage approaches. Our proposed CLIP-Q method (Compression Learning by In-Parallel Pruning-Quantization) compresses AlexNet by 51-fold, GoogLeNet by 10-fold, and ResNet-50 by 15-fold, while preserving the uncompressed network accuracies on ImageNet.

This work addresses the novel problem of one-shot one-class classification. The goal is to estimate a classification decision boundary for a novel class based on a single image example. Our method exploits transfer learning to model the transformation from a representation of the input, extracted by a Convolutional Neural Network, to a classification decision boundary. We use a deep neural network to learn this transformation from a large labelled dataset of images and their associated class decision boundaries generated from ImageNet, and then apply the learned decision boundary to classify subsequent query images. We tested our approach on several benchmark datasets and significantly outperformed the baseline methods.

Recently, there has been a paradigm shift in stereo matching with learning-based methods achieving the best results on all popular benchmarks. The success of these methods is due to the availability of training data with ground truth; training learning-based systems on these datasets has allowed them to surpass the accuracy of conventional approaches based on heuristics and assumptions. Many of these assumptions, however, had been validated extensively and hold for the majority of possible inputs. In this paper, we generate a matching volume leveraging both data with ground truth and conventional wisdom. We accomplish this by coalescing diverse evidence from a bidirectional matching process via random forest classifiers. We show that the resulting matching volume estimation method achieves similar accuracy to purely data-driven alternatives on benchmarks and that it generalizes to unseen data much better. In fact, the results we submitted to the KITTI benchmarks were generated using a classifier trained on the Middlebury dataset.

We present a technique for jointly denoising bursts of images taken from a handheld camera. In particular, we propose a convolutional neural network architecture for predicting spatially varying kernels that can both align and denoise frames, a synthetic data generation approach based on a realistic noise formation model, and an optimization guided by an annealed loss function to avoid undesirable local minima. Our model matches or outperforms the state-of-the-art across a wide range of noise levels on both real and synthetic data.

Using deep learning, this paper addresses the problem of joint object boundary detection and boundary motion estimation in videos, which we named boundary flow estimation. Boundary flow is an important mid-level visual cue as boundaries characterize objects' spatial extents, and the flow indicates objects' motions and interactions. Yet, most prior work on motion estimation has focused on dense object motion or feature points that may not necessarily reside on boundaries. For boundary flow estimation, we specify a new fully convolutional Siamese network (FCSN) that jointly estimates object-level boundaries in two consecutive frames. Boundary correspondences in the two frames are predicted by the same FCSN with a new, unconventional deconvolution approach. Finally, the boundary flow estimate is improved with an edgelet-based filtering. Evaluation is conducted on three tasks: boundary detection in videos, boundary flow estimation, and optical flow estimation. On boundary detection, we achieve the state-of-the-art performance on the benchmark VSB100 dataset. On boundary flow estimation, we present the first results on the Sintel training dataset. For optical flow estimation, we run the recent approach CPM-Flow but on the augmented input with our boundary-flow matches, and achieve significant performance improvement on the Sintel benchmark.

Top-down visual attention mechanisms have been used extensively in image captioning and visual question answering (VQA) to enable deeper image understanding through fine-grained analysis and even multiple steps of reasoning. In this work, we propose a combined bottom-up and top-down attention mechanism that enables attention to be calculated at the level of objects and other salient image regions. This is the natural basis for attention to be considered. Within our approach, the bottom-up mechanism (based on Faster R-CNN) proposes image regions, each with an associated feature vector, while the top-down mechanism determines feature weightings. Applying this approach to image captioning, our results on the MSCOCO test server establish a new state-of-the-art for the task, achieving CIDEr / SPICE / BLEU-4 scores of 117.9, 21.5 and 36.9, respectively. Demonstrating the broad applicability of the method, applying the same approach to VQA we obtain first place in the 2017 VQA Challenge.

Fully supervised methods for semantic segmentation require pixel-level class masks to train, the creation of which are expensive in terms of manual labour and time. In this work, we focus on weak supervision, developing a method for training a high-quality pixel-level classifier for semantic segmentation, using only image-level class labels as the provided ground-truth. Our method is formulated as a two-stage approach in which we first aim to create accurate pixel-level masks for the training images via a bootstrapping process, and then use these now-accurately segmented images as a proxy ground-truth in a more standard supervised setting. The key driver for our work is that in the target dataset we typically have reliable ground-truth image-level labels, while data crawled from the web may have unreliable labels, but can be filtered to comprise only easy images to segment, therefore having reliable boundaries. These two forms of information are complementary and we use this observation to build a novel bi-directional transfer learning. This framework transfers knowledge between two domains, target domain and web domain, bootstrapping the performance of weakly supervised semantic segmentation. Conducting experiments on the popular benchmark dataset PASCAL VOC 2012 based on both a VGG16 network and on ResNet50, we reach state-of-the-art performance with scores of 60.2% IoU and 63.9% IoU respectively.

In self-supervised learning one trains a model to solve a so-called pretext task on a dataset without the need for human annotation. The main objective, however, is to transfer this model to a target domain and task. Currently, the most effective transfer strategy is fine-tuning, which restricts one to use the same model or parts thereof for both pretext and target tasks. In this paper, we present a novel framework for self-supervised learning that overcomes limitations in designing and comparing different tasks, models, and data domains. In particular, our framework decouples the structure of the self-supervised model from the final task-specific fine-tuned model. This allows us to: 1) quantitatively assess previously incompatible models including handcrafted features; 2) show that deeper neural network models can learn better representations from the same pretext task; 3) transfer knowledge learned with a deep model to a shallower one and thus boost its learning. We use this framework to design a novel self-supervised task, which achieves state-of-the-art performance on the common benchmarks in PASCAL VOC 2007, ILSVRC12 and Places by a significant margin. A surprising result is that our learned features shrink the mAP gap between models trained via self-supervised learning and supervised learning from $5.9$ to $2.6$ in object detection on PASCAL VOC 2007.

Current Domain Adaptation (DA) methods based on deep architectures assume that the source samples arise from a single distribution. However, in practice most datasets can be regarded as mixtures of multiple domains. In these cases exploiting single-source DA methods for learning target classifiers may lead to sub-optimal, if not poor, results. In addition, in many applications it is difficult to manually provide the domain labels for all source data points, i.e. latent domains should be automatically discovered. This paper introduces a novel Convolutional Neural Network (CNN) architecture which (i) automatically discovers latent domains in visual datasets and (ii) exploits this information to learn robust target classifiers. Our approach is based on the introduction of two main components, which can be embedded into any existing CNN architecture: (i) a side branch that automatically computes the assignment of a source sample to a latent domain and (ii) novel layers that exploit domain membership information to appropriately align the distribution of the CNN internal feature representations to a reference distribution. We test our approach on publicly-available datasets, showing that it outperforms state-of-the-art multi-source DA methods by a large margin.

Deep neural networks are vulnerable to adversarial examples, which poses security concerns on these algorithms due to the potentially severe consequences. Adversarial attacks serve as an important surrogate to evaluate the robustness of deep learning models before they are deployed. However, most of existing adversarial attacks can only fool a black-box model with a low success rate. To address this issue, we propose a broad class of momentum-based iterative algorithms to boost adversarial attacks. By integrating the momentum term into the iterative process for attacks, our methods can stabilize update directions and escape from poor local maxima during the iterations, resulting in more transferable adversarial examples. To further improve the success rates for black-box attacks, we apply momentum iterative algorithms to an ensemble of models, and show that the adversarially trained models with a strong defense ability are also vulnerable to our black-box attacks. We hope that the proposed methods will serve as a benchmark for evaluating the robustness of various deep models and defense methods. With this method, we won the first places in NIPS 2017 Non-targeted Adversarial Attack and Targeted Adversarial Attack competitions.

Very deep convolutional neural networks offer excellent recognition results, yet their computational expense limits their impact for many real-world applications. We introduce BlockDrop, an approach that learns to dynamically choose which layers of a deep network to execute during inference so as to best reduce total computation without degrading prediction accuracy. Exploiting the robustness of Residual Networks (ResNets) to layer dropping, our framework selects on-the-fly which residual blocks to evaluate for a given novel image. In particular, given a pretrained ResNet, we train a policy network in an associative reinforcement learning setting for the dual reward of utilizing a minimal number of blocks while preserving recognition accuracy. We conduct extensive experiments on CIFAR and ImageNet. The results provide strong quantitative and qualitative evidence that these learned policies not only accelerate inference but also encode meaningful visual information. Built upon a ResNet-101 model, our method achieves a speedup of 20% on average, going as high as 36% for some images, while maintaining the same 76.4% top-1 accuracy on ImageNet.

A key problem in blind image quality assessment (BIQA) is how to effectively model the properties of human visual system in a data-driven manner. In this paper, we propose a simple and efficient BIQA model based on a novel framework which consists of a fully convolutional neural network (FCNN) and a pooling network to solve this problem. In principle, FCNN is capable of predicting a pixel-by-pixel similar quality map only from a distorted image by using the intermediate similarity maps derived from conventional full-reference image quality assessment methods. The predicted pixel-by-pixel quality maps have good consistency with the distortion correlations between the reference and distorted images. Finally, a deep pooling network regresses the quality map into a score. Experiments have demonstrated that our predictions outperform many state-of-the-art BIQA methods.