CVPR 2020 Papers

We present an approach for estimating surface normals from in-the-wild color images of faces. While data-driven strategies have been proposed for single face images, limited available ground truth data makes this problem difficult. To alleviate this issue, we propose a method that can leverage all available image and normal data, whether paired or not, thanks to a novel cross-modal learning architecture. In particular, we enable additional training with single modality data, either color or normal, by using two encoder-decoder networks with a shared latent space. The proposed architecture also enables face details to be transferred between the image and normal domains, given paired data, through skip connections between the image encoder and normal decoder. Core to our approach is a novel module that we call deactivable skip connections, which allows integrating both the auto-encoded and image-to-normal branches within the same architecture that can be trained end-to-end. This allows learning of a rich latent space that can accurately capture the normal information. We compare against state-of-the-art methods and show that our approach can achieve significant improvements, both quantitative and qualitative, with natural face images.

Text-based person search has drawn increasing attention due to its wide applications in video surveillance. However, most of the existing models depend heavily on paired image-text data, which is very expensive to acquire. Moreover, they always face huge performance drop when directly exploiting them to new domains. To overcome this problem, we make the first attempt to adapt the model to new target domains in the absence of pairwise labels, which combines the challenges from both cross-modal (text-based) person search and cross-domain person search. Specially, we propose a moment alignment network (MAN) to solve the cross-modal cross-domain person search task in this paper. The idea is to learn three effective moment alignments including domain alignment (DA), cross-modal alignment (CA) and exemplar alignment (EA), which together can learn domain-invariant and semantic aligned cross-modal representations to improve model generalization. Extensive experiments are conducted on CUHK Person Description dataset (CUHK-PEDES) and Richly Annotated Pedestrian dataset (RAP). Experimental results show that our proposed model achieves the state-of-the-art performances on five transfer tasks.

Exploiting photo-realistic synthetic data to train semantic segmentation models has received increasing attention over the past years. However, the domain mismatch between synthetic and real images will cause a significant performance drop when the model trained with synthetic images is directly applied to real-world scenarios. In this paper, we propose a new domain adaptation approach, called Pivot Interaction Transfer (PIT). Our method mainly focuses on constructing pivot information that is common knowledge shared across domains as a bridge to promote the adaptation of semantic segmentation model from synthetic domains to real-world domains. Specifically, we first infer the image-level category information about the target images, which is then utilized to facilitate pixel-level transfer for semantic segmentation, with the assumption that the interactive relation between the image-level category information and the pixel-level semantic information is invariant across domains. To this end, we propose a novel multi-level region expansion mechanism that aligns both the image-level and pixel-level information. Comprehensive experiments on the adaptation from both GTAV and SYNTHIA to Cityscapes clearly demonstrate the superiority of our method.

Face presentation attack detection (PAD) has been an urgent problem to be solved in the face recognition systems. Conventional approaches usually assume the testing and training are within the same domain; as a result, they may not generalize well into unseen scenarios because the representations learned for PAD may overfit to the subjects in the training set. In light of this, we propose an efficient disentangled representation learning for cross-domain face PAD. Our approach consists of disentangled representation learning (DR-Net) and multi-domain learning (MD-Net). DR-Net learns a pair of encoders via generative models that can disentangle PAD informative features from subject discriminative features. The disentangled features from different domains are fed to MD-Net which learns domain-independent features for the final cross-domain face PAD task. Extensive experiments on several public datasets validate the effectiveness of the proposed approach for cross-domain PAD.

Decomposing images of document pages into high-level semantic regions (e.g., figures, tables, paragraphs), document object detection (DOD) is fundamental for downstream tasks like intelligent document editing and understanding. DOD remains a challenging problem as document objects vary significantly in layout, size, aspect ratio, texture, etc. An additional challenge arises in practice because large labeled training datasets are only available for domains that differ from the target domain. We investigate cross-domain DOD, where the goal is to learn a detector for the target domain using labeled data from the source domain and only unlabeled data from the target domain. Documents from the two domains may vary significantly in layout, language, and genre. We establish a benchmark suite consisting of different types of PDF document datasets that can be utilized for cross-domain DOD model training and evaluation. For each dataset, we provide the page images, bounding box annotations, PDF files, and the rendering layers extracted from the PDF files. Moreover, we propose a novel cross-domain DOD model which builds upon the standard detection model and addresses domain shifts by incorporating three novel alignment modules: Feature Pyramid Alignment (FPA) module, Region Alignment (RA) module and Rendering Layer alignment (RLA) module. Extensive experiments on the benchmark suite substantiate the efficacy of the three proposed modules and the proposed method significantly outperforms the baseline methods. The project page is at https://github.com/kailigo/cddod.

Applying the knowledge of an object detector trained on a specific domain directly onto a new domain is risky, as the gap between two domains can severely degrade model's performance. Furthermore, since different instances commonly embody distinct modal information in object detection scenario, the feature alignment of source and target domain is hard to be realized. To mitigate these problems, we propose a Graph-induced Prototype Alignment (GPA) framework to seek for category-level domain alignment via elaborate prototype representations. In the nutshell, more precise instance-level features are obtained through graph-based information propagation among region proposals, and, on such basis, the prototype representation of each class is derived for category-level domain alignment. In addition, in order to alleviate the negative effect of class-imbalance on domain adaptation, we design a Class-reweighted Contrastive Loss to harmonize the adaptation training process. Combining with Faster R-CNN, the proposed framework conducts feature alignment in a two-stage manner. Comprehensive results on various cross-domain detection tasks demonstrate that our approach outperforms existing methods with a remarkable margin. Our code is available at https://github.com/ChrisAllenMing/GPA-detection.

We present a general framework for exemplar-based image translation, which synthesizes a photo-realistic image from the input in a distinct domain (e.g., semantic segmentation mask, or edge map, or pose keypoints), given an exemplar image. The output has the style (e.g., color, texture) in consistency with the semantically corresponding objects in the exemplar. We propose to jointly learn the cross-domain correspondence and the image translation, where both tasks facilitate each other and thus can be learned with weak supervision. The images from distinct domains are first aligned to an intermediate domain where dense correspondence is established. Then, the network synthesizes images based on the appearance of semantically corresponding patches in the exemplar. We demonstrate the effectiveness of our approach in several image translation tasks. Our method is superior to state-of-the-art methods in terms of image quality significantly, with the image style faithful to the exemplar with semantic consistency. Moreover, we show the utility of our method for several applications.

Mining informative negative instances are of central importance to deep metric learning (DML). However, the hard-mining ability of existing DML methods is intrinsically limited by mini-batch training, where only a mini-batch of instances are accessible at each iteration. In this paper, we identify a "slow drift" phenomena by observing that the embedding features drift exceptionally slow even as the model parameters are updating throughout the training process. It suggests that the features of instances computed at preceding iterations can considerably approximate to their features extracted by current model. We propose a cross-batch memory (XBM) mechanism that memorizes the embeddings of past iterations, allowing the model to collect sufficient hard negative pairs across multiple mini-batches - even over the whole dataset. Our XBM can be directly integrated into general pair-based DML framework.We demonstrate that, without bells and whistles, XBM augmented DML can boost the performance considerably on image retrieval. In particular, with XBM, a simple contrastive loss can have large R@1 improvements of 12%-22.5% on three large-scale datasets, easily surpassing the most sophisticated state-of-the-art methods [38, 27, 2], by a large margin. Our XBM is conceptually simple, easy to implement - using several lines of codes, and is memory efficient - with a negligible 0.2 GB extra GPU memory.

In recent years, cross-modal hashing (CMH) has attracted increasing attentions, mainly because its potential ability of mapping contents from different modalities, especially in vision and language, into the same space, so that it becomes efficient in cross-modal data retrieval. There are two main frameworks for CMH, differing from each other in whether semantic supervision is required. Compared to the unsupervised methods, the supervised methods often enjoy more accurate results, but require much heavier labors in data annotation. In this paper, we propose a novel approach that enables guiding a supervised method using outputs produced by an unsupervised method. Specifically, we make use of teacher-student optimization for propagating knowledge. Experiments are performed on two popular CMH benchmarks, i.e., the MIRFlickr and NUS-WIDE datasets. Our approach outperforms all existing unsupervised methods by a large margin.

We present CoverNet, a new method for multimodal, probabilistic trajectory prediction for urban driving. Previous work has employed a variety of methods, including multimodal regression, occupancy maps, and 1-step stochastic policies. We instead frame the trajectory prediction problem as classification over a diverse set of trajectories. The size of this set remains manageable due to the limited number of distinct actions that can be taken over a reasonable prediction horizon. We structure the trajectory set to a) ensure a desired level of coverage of the state space, and b) eliminate physically impossible trajectories. By dynamically generating trajectory sets based on the agent's current state, we can further improve our method's efficiency. We demonstrate our approach on public, real world self-driving datasets, and show that it outperforms state-of-the-art methods.

We present an approach for estimating the period with which an action is repeated in a video. The crux of the approach lies in constraining the period prediction module to use temporal self-similarity as an intermediate representation bottleneck that allows generalization to unseen repetitions in videos in the wild. We train this model, called RepNet, with a synthetic dataset that is generated from a large unlabeled video collection by sampling short clips of varying lengths and repeating them with different periods and counts. This combination of synthetic data and a powerful yet constrained model, allows us to predict periods in a class-agnostic fashion. Our model substantially exceeds the state of the art performance on existing periodicity (PERTUBE) and repetition counting (QUVA) benchmarks. We also collect a new challenging dataset called Countix ( 90 times larger than existing datasets) which captures the challenges of repetition counting in real-world videos. Project webpage: https://sites.google.com/view/repnet .

The ongoing success of visual question answering methods has been somwehat surprising given that, at its most general, the problem requires understanding the entire variety of both visual and language stimuli. It is particularly remarkable that this success has been achieved on the basis of comparatively small datasets, given the scale of the problem. One explanation is that this has been accomplished partly by exploiting bias in the datasets rather than developing deeper multi-modal reasoning. This fundamentally limits the generalization of the method, and thus its practical applicability. We propose a method that addresses this problem by introducing counterfactuals in the training. In doing so we leverage structural causal models for counterfactual evaluation to formulate alternatives, for instance, questions that could be asked of the same image set. We show that simulating plausible alternative training data through this process results in better generalization.

Despite Visual Question Answering (VQA) has realized impressive progress over the last few years, today's VQA models tend to capture superficial linguistic correlations in the train set and fail to generalize to the test set with different QA distributions. To reduce the language biases, several recent works introduce an auxiliary question-only model to regularize the training of targeted VQA model, and achieve dominating performance on VQA-CP. However, since the complexity of design, current methods are unable to equip the ensemble-based models with two indispensable characteristics of an ideal VQA model: 1) visual-explainable: the model should rely on the right visual regions when making decisions. 2) question-sensitive: the model should be sensitive to the linguistic variations in question. To this end, we propose a model-agnostic Counterfactual Samples Synthesizing (CSS) training scheme. The CSS generates numerous counterfactual training samples by masking critical objects in images or words in questions, and assigning different ground-truth answers. After training with the complementary samples (ie, the original and generated samples), the VQA models are forced to focus on all critical objects and words, which significantly improves both visual-explainable and question-sensitive abilities. In return, the performance of these models is further boosted. Extensive ablations have shown the effectiveness of CSS. Particularly, by building on top of the model LMH, we achieve a record-breaking performance of 58.95% on VQA-CP v2, with 6.5% gains.

We propose a cost volume-based neural network for depth inference from multi-view images. We demonstrate that building a cost volume pyramid in a coarse-to-fine manner instead of constructing a cost volume at a fixed resolution leads to a compact, lightweight network and allows us inferring high resolution depth maps to achieve better reconstruction results. To this end, we first build a cost volume based on uniform sampling of fronto-parallel planes across the entire depth range at the coarsest resolution of an image. Then, given current depth estimate, we construct new cost volumes iteratively on the pixelwise depth residual to perform depth map refinement. While sharing similar insight with Point-MVSNet as predicting and refining depth iteratively, we show that working on cost volume pyramid can lead to a more compact, yet efficient network structure compared with the Point-MVSNet on 3D points. We further provide detailed analyses of the relation between (residual) depth sampling and image resolution, which serves as a principle for building compact cost volume pyramid. Experimental results on benchmark datasets show that our model can perform 6x faster and has similar performance as state-of-the-art methods. Code is available at https://github.com/JiayuYANG/CVP-MVSNet

We present a method to infer physical material parameters, and even external boundaries, from the scanned motion of a homogeneous deformable object via the solution of an inverse problem. Parameters are estimated from real-world data sources such as sparse observations from a Kinect sensor without correspondences. We introduce a novel Lagrangian-Eulerian optimization formulation, including a cost function that penalizes differences to observations during an optimization run. This formulation matches correspondence-free, sparse observations from a single-view depth image with a finite element simulation of deformable bodies. In a number of tests using synthetic datasets and real-world measurements, we analyse the robustness of our approach and the convergence behavior of the numerical optimization scheme.

In this paper, we tackle the task of establishing dense visual correspondences between images containing objects of the same category. This is a challenging task due to large intra-class variations and a lack of dense pixel level annotations. We propose a convolutional neural network architecture, called adaptive neighbourhood consensus network (ANC-Net), that can be trained end-to-end with sparse key-point annotations, to handle this challenge. At the core of ANC-Net is our proposed non-isotropic 4D convolution kernel, which forms the building block for the adaptive neighbourhood consensus module for robust matching. We also introduce a simple and efficient multi-scale self-similarity module in ANC-Net to make the learned feature robust to intra-class variations. Furthermore, we propose a novel orthogonal loss that can enforce the one-to-one matching constraint. We thoroughly evaluate the effectiveness of our method on various benchmarks, where it substantially outperforms state-of-the-art methods.

Accurate bounding box estimation has recently attracted much attention in the tracking community because traditional multi-scale search strategies cannot estimate tight bounding boxes in many challenging scenarios involving changes to the target. A tracker capable of detecting target corners can flexibly adapt to such changes, but existing corner detection based tracking methods have not achieved adequate success. We analyze the reasons for their failure and propose a state-of-the-art tracker that performs correlation-guided attentional corner detection in two stages. First, a region of interest (RoI) is obtained by employing an efficient Siamese network to distinguish the target from the background. Second, a pixel-wise correlation-guided spatial attention module and a channel-wise correlation-guided channel attention module exploit the relationship between the target template and the RoI to highlight corner regions and enhance features of the RoI for corner detection. The correlation-guided attention modules improve the accuracy of corner detection, thus enabling accurate bounding box estimation. When trained on large-scale datasets using a novel RoI augmentation strategy, the performance of the proposed tracker, running at a high speed of 70 FPS, is comparable with that of state-of-the-art trackers in meeting five challenging performance benchmarks.

According to existing studies, human body edge and pose are two beneficial factors to human parsing. The effectiveness of each of the high-level features (edge and pose) is confirmed through the concatenation of their features with the parsing features. Driven by the insights, this paper studies how human semantic boundaries and keypoint locations can jointly improve human parsing. Compared with the existing practice of feature concatenation, we find that uncovering the correlation among the three factors is a superior way of leveraging the pivotal contextual cues provided by edges and poses. To capture such correlations, we propose a Correlation Parsing Machine (CorrPM) employing a heterogeneous non-local block to discover the spatial affinity among feature maps from the edge, pose and parsing. The proposed CorrPM allows us to report new state-of-the-art accuracy on three human parsing datasets. Importantly, comparative studies confirm the advantages of feature correlation over the concatenation.

The single image super-resolution task is one of the most examined inverse problems in the past decade. In the recent years, Deep Neural Networks (DNNs) have shown superior performance over alternative methods when the acquisition process uses a fixed known downscaling kernel---typically a bicubic kernel. However, several recent works have shown that in practical scenarios, where the test data mismatch the training data (e.g. when the downscaling kernel is not the bicubic kernel or is not available at training), the leading DNN methods suffer from a huge performance drop. Inspired by the literature on generalized sampling, in this work we propose a method for improving the performance of DNNs that have been trained with a fixed kernel on observations acquired by other kernels. For a known kernel, we design a closed-form correction filter that modifies the low-resolution image to match one which is obtained by another kernel (e.g. bicubic), and thus improves the results of existing pre-trained DNNs. For an unknown kernel, we extend this idea and propose an algorithm for blind estimation of the required correction filter. We show that our approach outperforms other super-resolution methods, which are designed for general downscaling kernels.