WACV 2020 Papers

Hashing has attracted attention in recent years due to the rapid growth of image and video data on the web. Benefiting from recent advances in deep learning, deep supervised hashing has achieved promising results for image retrieval. However, existing methods are either less efficient in data usage or incapable of learning linearly discriminative binary codes. In this paper, we revisit linear discriminative analysis and propose a linear discriminative hashing (LDH) objective that is efficient in training and achieves better accuracy in retrieval. With the joint supervision of a classification loss, we design a robust deep network to obtain binary codes that are inter-class separable and intra-class compact, which provides better representations for image retrieval. We conduct extensive experiments on three benchmark datasets, and our LDH algorithm performs favorably against existing state-of-the-art deep supervised hashing methods.

Despite the significant advances in recent years, Generative Adversarial Networks (GANs) are still notoriously hard to train. In this paper, we propose three novel curriculum learning strategies for training GANs. All strategies are first based on ranking the training images by their difficulty scores, which are estimated by a state-of-the-art image difficulty predictor. Our first strategy is to divide images into gradually more difficult batches. Our second strategy introduces a novel curriculum loss function for the discriminator that takes into account the difficulty scores of the real images. Our third strategy is based on sampling from an evolving distribution, which favors the easier images during the initial training stages and gradually converges to a uniform distribution, in which samples are equally likely, regardless of difficulty. We compare our curriculum learning strategies with the classic training procedure on two tasks: image generation and image translation. Our experiments indicate that all strategies provide faster convergence and superior results. For example, our best curriculum learning strategy applied on spectrally normalized GANs (SNGANs) fooled human annotators in thinking that generated CIFAR-like images are real in 25.0% of the presented cases, while the SNGANs trained using the classic procedure fooled the annotators in only 18.4% cases. Similarly, in image translation, the human annotators preferred the images produced by the Cycle-consistent GAN (CycleGAN) trained using curriculum learning in 40.5% cases and those produced by CycleGAN based on classic training in only 19.8% cases, 39.7% cases being labeled as ties.

Generating human videos based on single images entails the challenging simultaneous generation of realistic and visual appealing appearance and motion. In this context, we propose a novel conditional GAN architecture, namely ImaGINator, which given a single image, a condition (label of a facial expression or action) and noise, decomposes appearance and motion in both latent and high level feature spaces, generating realistic videos. This is achieved by (i)a novel spatio-temporal fusion scheme, which generates dynamic motion, while retaining appearance throughout the full video sequence by transmitting appearance (originating from the single image) through all layers of the network. In addition, we propose (ii) a novel transposed (1+2)D convolution, factorizing the transposed 3D convolutional filters into separate transposed temporal and spatial components, which yields significantly gains in video quality and speed. We extensively evaluate our approach on the facial expression datasets MUG and UvA-NEMO, as well as on the action datasets NATOPS and Weizmann. We show that our approach achieves significantly better quantitative and qualitative results than the state-of-the-art. The source code and models are available under https://github.com/wyhsirius/ImaGINator.

Image denoising methods must effectively model, implicitly or explicitly, the vast diversity of patterns and textures that occur in natural images. This is challenging, even for modern methods that leverage deep neural networks trained to regress to clean images from noisy inputs. One recourse is to rely on "internal" image statistics, by searching for similar patterns within the input image itself. In this work, we propose a new method for natural image denoising that trains a deep neural network to determine whether patches in a noisy image input share common underlying patterns. Given a pair of noisy patches, our network predicts whether different sub-band coefficients of the original noise-free patches are similar. The denoising algorithm then aggregates matched coefficients to obtain an initial estimate of the clean image. Finally, this estimate is provided as input, along with the original noisy image, to a standard regression-based denoising network. Experiments show that our method achieves state-of-the-art color image denoising performance, including with a blind version that trains a common model for a range of noise levels, and does not require knowledge of level of noise in an input image. Our approach also has a distinct advantage when training with limited amounts of training data.

This paper presents a generic face animator that can control the pose and expressions of a given face image. The animation is driven by human interpretable control signals consisting of head pose angles and the Action Unit (AU) values. The control information can be obtained from multiple sources including external driving videos and manual controls. Due to the interpretable nature of the driving signal, one can easily mix the information between multiple sources (e.g. pose from one image and expression from another) and apply selective post- production editing. The proposed face animator is implemented as a two-stage neural network model that is learned in a self-supervised manner using a large video collection. The proposed Interpretable and Controllable face reenactment network (ICface) is compared to the state-of-the-art neural network-based face animation techniques in multiple tasks. The results indicate that ICface produces better visual quality while being more versatile than most of the comparison methods. The introduced model could provide a lightweight and easy to use tool for a multitude of advanced image and video editing tasks. The program code will be publicly available upon the acceptance of the paper.

We introduce I-MOVE, the first publicly available RGB-D/stereo dataset for estimating velocities of independently moving objects. Velocity estimation given RGB-D data is an unsolved problem. The I-MOVE dataset provides an opportunity for generalizable velocity estimation models to be created and have their performance be accurately and fairly measured. The dataset features various outdoor and indoor scenes of single and multiple moving objects. Compared to other datasets, I-MOVE is unique because the RGB-D data and speed for each object are supplied for a variety of different settings/environments, objects, and motions. The dataset includes training and test sequences captured from four different depth camera views and three 4K-stereo setups. The data are also time-synchronized with three Doppler radars to provide the magnitude of velocity ground truth. The I-MOVE dataset includes complex scenes from moving pedestrians via walking and biking to multiple rolling objects, all captured with the seven cameras, providing over 500 Depth/Stereo videos.

Neural style transfer has been a "killer app" for deep learning, drawing attention from and advertising the effectiveness to both the academic and the general public. However, we have found by ablative experiments that optimizing an image in the way neural style transfer does, while the objective functions (or more precisely, the functions to transform raw images to corresponding feature maps being compared) are constructed without pretrained weights or biases, worked al- most as well. We can even factor out the deepness (multiple layers of alternating linear and nonlinear transformations) alltogether and have neural style transfer work to a certain extent. This raises the question how much of the the current success of deep learning in computer vision should be attributed to training, structure or simply spatially aggregating the image.

We propose to predict histograms of object sizes in crowded scenes directly without any explicit object instance segmentation. What makes this task challenging is the high density of objects (of the same category), which makes instance identification hard. Instead of explicitly segmenting object instances, we show that directly learning histograms of object sizes improves accuracy while using drastically less parameters. This is very useful for application scenarios where explicit, pixel-accurate instance segmentation is not needed, but their lies interest in the overall distribution of instance sizes. Our core applications are in biology, where we estimate the size distribution of soldier fly larvae, and medicine, where we estimate the size distribution of cancer cells as an intermediate step to calculate tumor cellularity score. Given an image with hundreds of small object instances, we output the total count and the size histogram. We also provide a new data set for this task, the FlyLarvae data set, which consists of 11,000 larvae instances labeled pixel-wise. Our method results in an overall improvement in the count and size distribution prediction as compared to state-of-the-art instance segmentation method Mask R-CNN.

A video super-resolution technique is expected to generate a `sharp' upsampled video. The sharpness in the generated video comes from the precise prediction of the high-frequency details (e.g. object edges). Thus high-frequency prediction becomes a vital sub-problem of the super-resolution task. To generate a sharp-upsampled video, this paper proposes an upsampling network architecture `HFR-Net' that works on the principle of `explicit refinement and fusion of high-frequency details'. To implement this principle and to train HFR-Net, a novel technique named 2-phase progressive-retrogressive training is being proposed. Additionally, a method called dual motion warping is also being introduced to preprocess the videos that have varying motion intensities (slow and fast). Results on multiple video datasets demonstrate the improved performance of our approach over the current state-of-the-art.

What's the most accurate 3D model of your face you can obtain while sitting at your desk? We attempt to answer this question in our work. High fidelity face reconstructions have so far been limited to either studio settings or through expensive 3D scanners. On the other hand, unconstrained reconstruction methods are typically limited by low-capacity models. Our method reconstructs accurate face geometry of a subject using a video shot from a smartphone in an unconstrained environment. Our approach takes advantage of recent advances in visual SLAM, keypoint detection, and object detection to improve accuracy and robustness. By not being constrained to a model subspace, our reconstructed meshes capture important details while being robust to noise and being topologically consistent. Our evaluations show that our method outperforms current single and multi-view baselines by a significant margin, both in terms of geometric accuracy and in capturing person-specific details important for making realistic looking models.

Egocentric vision holds great promises for increasing access to visual information and improving the quality of life for people with visual impairments, with object recognition being one of the daily challenges for this population. While we strive to improve recognition performance, it remains difficult to identify which object is of interest to the user; the object may not even be included in the frame due to challenges in camera aiming without visual feedback. Also, gaze information, commonly used to infer the area of interest in egocentric vision, is often not dependable. However, blind users often tend to include their hand either interacting with the object that they wish to recognize or simply placing it in proximity for better camera aiming. We propose localization models that leverage the presence of the hand as the contextual information for priming the center area of the object of interest. In our approach, hand segmentation is fed to either the entire localization network or its last convolutional layers. Using egocentric datasets from sighted and blind individuals, we show that the hand-priming achieves higher precision than other approaches, such as fine-tuning, multi-class, and multi-task learning, which also encode hand-object interactions in localization.

A graph neural network (GNN) for image understanding based on multiple cues is proposed in this paper. Compared to traditional feature and decision fusion approaches that neglect the fact that features can interact and exchange information, the proposed GNN is able to pass information among features extracted from different models. Two image understanding tasks, namely group-level emotion recognition (GER) and event recognition, which are highly semantic and require the interaction of several deep models to synthesize multiple cues, were selected to validate the performance of the proposed method. It is shown through experiments that the proposed method achieves state-of-the-art performance on the selected image understanding tasks. In addition, a new group-level emotion recognition database is introduced and shared in this paper.

Multiple object tracking (MOT) task requires reasoning the states of all targets and associating these targets in a global way. However, existing MOT methods mostly focus on the local relationship among objects and ignore the global relationship. Some methods formulate the MOT problem as a graph optimization problem. However, these methods are based on static graphs, which are seldom updated. To solve these problems, we design a new near-online MOT method with an end-to-end graph network. Specifically, we design an appearance graph network and a motion graph network to capture the appearance and the motion similarity separately. The updating mechanism is carefully designed in our graph network, which means that nodes, edges and the global variable in the graph can be updated. The global variable can capture the global relationship to help tracking. Finally, a strategy to handle missing detections is proposed to remedy the defect of the detectors. Our method is evaluated on both the MOT16 and the MOT17 benchmarks, and experimental results show the encouraging performance of our method.

The computer vision community is witnessing an unprecedented rate of new tasks being proposed and addressed, thanks to the deep convolutional networks' capability to find complex mappings from X to Y. The advent of each task often accompanies the release of a large-scale annotated dataset, for supervised training of deep network. However, it is expensive and time-consuming to manually label sufficient amount of training data. Therefore, it is important to develop algorithms that can leverage off-the-shelf labeled dataset to learn useful knowledge for the target task. While previous works mostly focus on transfer learning from a single source, we study multi-source transfer across domains and tasks (MS-DTT), in a semi-supervised setting. We propose GradMix, a model-agnostic method applicable to any model trained with gradient-based learning rule, to transfer knowledge via gradient descent by weighting and mixing the gradients from all sources during training. GradMix follows a meta-learning objective, which assigns layer-wise weights to the source gradients, such that the combined gradient follows the direction that minimize the loss for a small set of samples from the target dataset. In addition, we propose to adaptively adjust the learning rate for each mini-batch based on its importance to the target task, and a pseudo-labeling method to leverage the unlabeled samples in the target domain. We conduct MS-DTT experiments on two tasks: digit recognition and action recognition, and demonstrate the advantageous performance of the proposed method against multiple baselines.

Divide and Conquer is a well-established approach in the literature that has efficiently solved a variety of problems. However, it is yet to be explored in full in solving image super-resolution. To predict a sharp up-sampled image, this work proposes a divide and conquer approach based wide and deep network (WDN) that divides the 4x up-sampling problem into 32 disjoint subproblems that can be solved simultaneously and independently of each other. Half of these subproblems deal with predicting the overall features of the high-resolution image, while the remaining are exclusively for predicting the finer details. Additionally, a technique that is found to be more effective in calibrating the pixel intensities has been proposed. Results obtained on multiple datasets demonstrate the improved performance of the proposed wide and deep network over state-of-the-art methods.

We present an alternative to the paradigm of density regression widely being employed for tackling crowd counting. In the prevalent regression approach, a model is trained for mapping images to its crowd density rather than counting by detecting every person. This framework is motivated from the difficulty to discriminate humans in highly dense crowds where unfavorable perspective, occlusion and clutter are prevalent. Though regression methods estimate overall crowd counts pretty well, localization of individual persons suffers and varies considerably across the entire density spectrum. Moreover, individual detection of people aids more explainable practical systems than predicting blind crowd count or density map. Hence, we move away from density regression and reformulate the task as localized dot prediction in dense crowds. Our dot detection model, DD-CNN, is trained for pixel-wise binary classification to detect people instead of regressing local crowd density. In order to handle severe scale variation and detect people of all scales with accurate dots, we use a novel multi-scale architecture which does not require any ground truth scale information. This training regime, which incorporates top-down feedback, helps our model to localize people in sparse as well as dense crowds. Our model delivers superior counting performance on major crowd datasets. We also evaluate on some additional metrics and evidence superior localization of the dot detection formulation.

Global contextual dependency is important for semantic segmentation of 3D point clouds. However, most existing approaches stack feature extraction layers to enlarge the receptive field to aggregate more contextual information of points along the spatial dimension. In this paper, we propose a Point Global Context Reasoning (PointGCR) module to capture global contextual information along the channel dimension. In PointGCR, an undirected graph representation (namely, ChannelGraph) is used to learn channel independencies. Specifically, channel maps are first represented as graph nodes and the independencies between nodes are then represented as graph edges. PointGCR is a plug-andplay and end-to-end trainable module. It can easily be integrated into an existing segmentation network and achieves a significant performance improvement. We conduct extensive experiments to evaluate the proposed PointGCR module on both indoor and outdoor datasets. Experimental results show that our PointGCR module efficiently captures global contextual dependencies and significantly improve the segmentation performance of several existing networks.

Skeleton-based action recognition has attracted more and more attention in recent years. Besides, the rapid development of deep learning has greatly improved the performance.However, the current exploration of action cooccurrence is still not comprehensive enough. Most existing works only mine co-occurrence features from the temporal or spatial domain seperately, and it's common to combine them in the end. Different from previous works, our approach is able to learn temporal and spatial co-occurrence features integratedly and globally, which is called spatio-temporal-unit feature enhancement (STUFE). In order to better align the skeleton data, we introduce a novel method for skeleton data preprocessing called active coordinate system conversion (ACSC). A coordinate system can be learned automatically to transform skeleton samples for alignment. By the way, the proposed methods are compatible with current two types of mainstream models, the CNN-based and GCN-based models. Finally, on the two benchmarks of NTU-RGB+D and SBU Kinect Interaction, we validated our methods based on two mainstream models.The results show that our methods achieve the state-of-the-art.

This paper addresses the problem of determining dense pixel correspondences between two images and its application to geometric correspondence verification in image retrieval. The main contribution is a geometric correspondence verification approach for re-ranking a shortlist of retrieved database images based on their dense pair-wise matching with the query image at a pixel level. We determine a set of cyclically consistent dense pixel matches between the pair of images and evaluate local similarity of matched pixels using neural network based image descriptors. Final re-ranking is based on a novel similarity function, which fuses the local similarity metric with a global similarity metric and a geometric consistency measure computed for the matched pixels. For dense matching our approach utilizes a modified version of a recently proposed dense geometric correspondence network (DGC-Net), which we also improve by optimizing the architecture. The proposed model and similarity metric compare favourably to the state-of-the-art image retrieval methods. In addition, we apply our method to the problem of long-term visual localization demonstrating promising results and generalization across datasets.