WACV 2022 Papers

We focus on the problem of novel-view human action synthesis. Given an action video, the goal is to generate the same action from an unseen viewpoint. Naturally, novel view video synthesis is more challenging than image synthesis. It requires the synthesis of a sequence of realistic frames with temporal coherency. Besides, transferring the different actions to a novel target view requires awareness of action category and viewpoint change simultaneously. To address these challenges we propose a novel framework named Pose-guided Action Separable Generative Adversarial Net (PAS-GAN), which utilizes pose to alleviate the difficulty of this task. First, we propose a recurrent pose-transformation module which transforms actions from the source view to the target view and generates novel view pose sequence in 2D coordinate space. Second, a well-transformed pose sequence enables us to separatethe action and background in the target view. We employ a novel local-global spatial transformation module to effectively generate sequential video features in the target view using these action and background features. Finally, the generated video features are used to synthesize human action with the help of a 3D decoder. Moreover, to focus on dynamic action in the video, we propose a novel multi-scale action-separable loss which further improves the video quality. We conduct extensive experiments on two large-scale multi-view human action datasets, NTU-RGBD and PKU-MMD, demonstrating the effectiveness of PAS-GAN which outperforms existing approaches.

Recent progress on action recognition has mainly focused on RGB and optical flow features. In this paper, we approach the problem of joint-based action recognition. Unlike other modalities, constellation of joints and their motion generate models with succinct human motion information for activity recognition. We present a new model for joint-based action recognition, which first extracts motion features from each joint separately through a shared motion encoder before performing collective reasoning. Our joint selector module re-weights the joint information to select the most discriminative joints for the task. We also propose a novel joint-contrastive loss that pulls together groups of joint features which convey the same action. We strengthen the joint-based representations by using a geometry-aware data augmentation technique which jitters pose heatmaps while retaining the dynamics of the action. We show large improvements over the current state-of-the-art joint-based approaches on JHMDB, HMDB, Charades, AVA action recognition datasets. A late fusion with RGB and Flow-based approaches yields additional improvements. Our model also outperforms the existing baseline on Mimetics, a dataset with out-of-context actions.

In this paper, a real-time method called PoP-Net is proposed to predict multi-person 3D poses from a depth image. PoP-Net learns to predict bottom-up part representations and top-down global poses in a single shot. Specifically, a new part-level representation, called Truncated Part Displacement Field (TPDF), is introduced which enables an explicit fusion process to unify the advantages of bottom-up part detection and global pose detection. Meanwhile, an effective mode selection scheme is introduced to automatically resolve the conflicting cases between global pose and part detections. Finally, due to the lack of high-quality depth datasets for developing multi-person 3D pose estimation, we introduce Multi-Person 3D Human Pose Dataset (MP-3DHP) as a new benchmark. MP-3DHP is designed to enable effective multi-person and background data augmentation in model training, and to evaluate 3D human pose estimators under uncontrolled multi-person scenarios. We show that PoP-Net achieves the state-of-the-art results both on MP-3DHP and on the widely used ITOP dataset, and has significant advantages in efficiency for multi-person processing. MP-3DHP Dataset and the evaluation code have been made available at: https://github.com/oppo-us-research/PoP-Net.

Although recent semantic segmentation methods have made remarkable progress, they still rely on large amounts of annotated training data, which are often infeasible to collect in the autonomous driving scenario. Previous works usually tackle this issue with Unsupervised Domain Adaptation (UDA), which entails training a network on synthetic images and applying the model to real ones while minimizing the discrepancy between the two domains. Yet, these techniques do not consider additional information that may be obtained from other tasks. Differently, we propose to exploit self-supervised monocular depth estimation to improve UDA for semantic segmentation. On one hand, we deploy depth to realize a plug-in component which can inject complementary geometric cues into any existing UDA method. We further rely on depth to generate a large and varied set of samples to Self-Train the final model. Our whole proposal allows for achieving state-of-the-art performance (58.8 mIoU) in the GTA5->CS benchmark benchmark. Code is available at https://github.com/CVLAB-Unibo/d4-dbst.

In this paper we consider the task of semantic segmentation in autonomous driving applications. Specifically, we consider the cross-domain few-shot setting where training can use only few real-world annotated images and many annotated synthetic images. In this context, aligning the domains is made more challenging by the pixel-wise class imbalance that is intrinsic in the segmentation and that leads to ignoring the underrepresented classes and overfitting the well represented ones. We address this problem with a novel framework called Pixel-By-Pixel Cross-Domain Alignment (PixDA). We propose a novel pixel-by-pixel domain adversarial loss following three criteria: (i) align the source and the target domain for each pixel, (ii) avoid negative transfer on the correctly represented pixels, and (iii) regularize the training of infrequent classes to avoid overfitting. The pixel-wise adversarial training is assisted by a novel sample selection procedure, that handles the imbalance between source and target data, and a knowledge distillation strategy, that avoids overfitting towards the few target images. We demonstrate on standard synthetic-to-real benchmarks that PixDA outperforms previous state-of-the-art methods in (1-5)-shot settings.

Semi-supervised video object segmentation (VOS) aims to track a designated object present in the initial frame of a video at the pixel level. To fully exploit the appearance information of an object, pixel-level feature matching is widely used in VOS. Conventional feature matching runs in a surjective manner, i.e., only the best matches from the query frame to the reference frame are considered. Each location in the query frame refers to the optimal location in the reference frame regardless of how often each reference frame location is referenced. This works well in most cases and is robust against rapid appearance variations, but may cause critical errors when the query frame contains background distractors that look similar to the target object. To mitigate this concern, we introduce a bijective matching mechanism to find the best matches from the query frame to the reference frame and vice versa. Before finding the best matches for the query frame pixels, the optimal matches for the reference frame pixels are first considered to prevent each reference frame pixel from being overly referenced. As this mechanism operates in a strict manner, i.e., pixels are connected if and only if they are the sure matches for each other, it can effectively eliminate background distractors. In addition, we propose a mask embedding module to improve the existing mask propagation method. By utilizing multiple historic masks and their variations, it can effectively capture the position information of a target object.

Deep neural networks (DNNs) have become essential for processing the vast amounts of aerial imagery collected using earth-observing satellite platforms. However, DNNs are vulnerable towards adversarial examples, and it is expected that this weakness also plagues DNNs for aerial imagery. In this work, we demonstrate one of the first efforts on physical adversarial attacks on aerial imagery, whereby adversarial patches were optimised, fabricated and installed on or near target objects (cars) to significantly reduce the efficacy of an object detector applied on overhead images. Physical adversarial attacks on aerial images, particularly those captured from satellite platforms, are challenged by atmospheric factors (lighting, weather, seasons) and the distance between the observer and target. To investigate the effects of these challenges, we devised novel experiments and metrics to evaluate the efficacy of physical adversarial attacks against object detectors in aerial scenes. Our results indicate the palpable threat posed by physical adversarial attacks towards DNNs for processing satellite imagery.

Photorealistic style transfer is an image editing task with the goal to modify an image to match the style of another image while ensuring the result looks like a real photograph. A limitation of existing models is that they have many parameters, which in turn prevents their use for larger image resolutions and leads to slower run-times. We introduce two mechanisms that enable our design of a more compact model that we call PhotoWCT2, which preserves state-of-art stylization strength and photorealism. First, we introduce blockwise training to perform coarse-to-fine feature transformations that enable state-of-art stylization strength in a single autoencoder in place of the inefficient cascade of four autoencoders used in PhotoWCT. Second, we introduce skip connections of high-frequency residuals in order to preserve image quality when applying the sequential coarse-to-fine feature transformations. Our PhotoWCT2 model requires fewer parameters (e.g., 30.3% fewer) while supporting higher resolution images (e.g., 4K) and achieving faster stylization than existing models.

In this paper, we present a novel perceptual consistency perspective on video semantic segmentation, which can capture both temporal consistency and pixel-wise correctness. Given two nearby video frames, perceptual consistency measures how much the segmentation decisions agree with the pixel correspondences obtained via matching general perceptual features. More specifically, for each pixel in one frame, we find the most perceptually correlated pixel in the other frame. Our intuition is that such a pair of pixels are highly likely to belong to the same class. Next, we assess how much the segmentation agrees with such perceptual correspondences, based on which we derive the perceptual consistency of the segmentation maps across these two frames. Utilizing perceptual consistency, we can evaluate the temporal consistency of video segmentation by measuring the perceptual consistency over consecutive pairs of segmentation maps in a video. Furthermore, given a sparsely labeled test video, perceptual consistency can be utilized to aid with predicting the pixel-wise correctness of the segmentation on an unlabeled frame. More specifically, by measuring the perceptual consistency between the predicted segmentation and the available ground truth on a nearby frame and combining it with the segmentation confidence, we can accurately assess the classification correctness on each pixel. Our experiments show that the proposed perceptual consistency can more accurately evaluate the temporal consistency of video segmentation as compared to flow-based measures. Furthermore, it can help more confidently predict segmentation accuracy on unlabeled test frames, as compared to using classification confidence alone. Finally, our proposed measure can be used as a regularizer during the training of segmentation models, which leads to more temporally consistent video segmentation while maintaining accuracy.

This paper proposes a method for automatically recovering data from chart images. In particular we focus on the task of estimating line charts, as the most common chart type, in a fully automatic way that handles line occlusions, as well as lines of different styles, e.g., dashed or dotted. For this, we first train a single semantic segmentation network to predict probability maps for each different line styles. We then construct a graph based on this output and formulate the line tracing task as a minimum-cost-flow problem, optimizing a cost function using linear programming. From the traced lines, the axes, and text labels, we recover the numerical values used to generate the chart. In experiments on six datasets, containing both synthesized and crawled images, we show significant improvements over prior work.

Modern AI tools, such as generative adversarial networks, have transformed our ability to create and modify visual data with photorealistic results. However, one of the deleterious side-effects of these advances is the emergence of nefarious uses in manipulating information in visual data, such as through the use of deep fakes. We propose a novel architecture for preserving the provenance of semantic information in images to make them less susceptible to deep fake attacks. Our architecture includes semantic signing and verification steps. We apply this architecture to verifying two types of semantic information: individual identities (faces) and whether the photo was taken indoors or outdoors. Verification in both cases carefully accounts for a collection of common image transformation, such as translation, scaling, cropping, and small rotations, and rejects adversarial transformations, such as adversarially perturbed or, in the case of face verification, swapped faces. Experiments demonstrate that in the case of provenance of faces in an image, our approach is robust to black-box adversarial transformations (which are rejected) as well as benign transformations (which are accepted), with few false negatives and false positives. Background verification, on the other hand, is susceptible to black-box adversarial examples, but becomes significantly more robust after adversarial training.

Collaborative training of neural networks leverages distributed data by exchanging gradient information between different clients. Although training data entirely resides with the clients, recent work shows that training data can be reconstructed from such exchanged gradient information. To enhance privacy, gradient perturbation techniques have been proposed. However, they come at the cost of reduced model performance, increased convergence time, or increased data demand. In this paper, we introduce PRECODE, a PRivacy EnhanCing mODulE that can be used as generic extension for arbitrary model architectures. We propose a simple yet effective realization of PRECODE using variational modeling. The stochastic sampling induced by variational modeling effectively prevents privacy leakage from gradients and in turn preserves privacy of data owners. We evaluate PRECODE using state of the art gradient inversion attacks on two different model architectures trained on three datasets. In contrast to commonly used defense mechanisms, we find that our proposed modification consistently reduces the attack success rate to 0% while having almost no negative impact on model training and final performance. As a result, PRECODE reveals a promising path towards privacy enhancing model extensions.

We push forward neural network compression research by exploiting a novel challenging task of large-scale conditional generative adversarial networks (GANs) compression. To this end, we propose a gradually shrinking GAN (PPCD-GAN) by introducing progressive pruning residual block (PP-Res) and class-aware distillation. The PP-Res is an extension of the conventional residual block where each convolutional layer is followed by a learnable mask layer to progressively prune network parameters as training proceeds. The class-aware distillation, on the other hand, enhances the stability of training by transferring immense knowledge from a well-trained teacher model through instructive attention maps. We train the pruning and distillation processes simultaneously on a well-known GAN architecture in an end-to-end manner. After training, all redundant parameters as well as the mask layers are discarded, yielding a lighter network while retaining the performance. We comprehensively illustrate, on ImageNet 128 x 128 dataset, PPCD-GAN reduces up to 5.2x (81%) parameters against state-of-the-arts while keeping better performance.

In this work, we focus on supervised domain adaptation for object detection in few-shot loose annotation setting, where the source images are sufficient and fully labeled but the target images are few-shot and loosely annotated. As annotated objects exist in the target domain, instance level alignment can be utilized to improve the performance. Traditional methods conduct the instance level alignment by semantically aligning the distributions of paired object features with domain adversarial training. Although it is demonstrated that point-wise surrogates of distribution alignment provide a more effective solution in few-shot classification tasks across domains, this point-wise alignment approach has not yet been extended to object detection. In this work, we propose a method that extends the point-wise alignment from classification to object detection. Moreover, in the few-shot loose annotation setting, the background ROIs of target domain suffer from severe label noise problem, which may make the point-wise alignment fail. To this end, we exploit moving average centroids to mitigate the label noise problem of background ROIs. Meanwhile, we exploit point-wise alignment over instances and centroids to tackle the problem of scarcity of labeled target instances. Hence this method is not only robust against label noises of background ROIs but also robust against the scarcity of labeled target objects. Experimental results show that the proposed instance level alignment method brings significant improvement compared with the baseline and is superior to state-of-the-art methods.

In recent years, many works in the video action recognition literature have shown that two stream models (combining spatial and temporal input streams) are necessary for achieving state-of-the-art performance. In this paper we show the benefits of including yet another stream based on human pose estimated from each frame --- specifically by rendering pose on input RGB frames. At first blush, this additional stream may seem redundant given that human pose is fully determined by RGB pixel values --- however we show (perhaps surprisingly) that this simple and flexible addition can provide complementary gains. Using this insight, we propose a new model, which we dub PERF-Net (short for Pose Empowered RGB-Flow Net), which combines this new pose stream with the standard RGB and flow based input streams via distillation techniques and show that our model outperforms the state-of-the-art by a large margin in a number of human action recognition datasets while not requiring flow or pose to be explicitly computed at inference time. The proposed pose stream is also part of the winner solution of the ActivityNet Kinetics Challenge 2020.

In few-shot classification, the primary goal is to learn representations from a few samples that generalize well for novel classes. In this paper, we propose an efficient low displacement rank (LDR) regularization strategy termed Ortho-Shot; a technique that imposes orthogonal regularization on the convolutional layers of a few-shot classifier, which is based on the doubly-block toeplitz (DBT) matrix structure. The regularized convolutional layers of the few-shot classifier enhances model generalization and intra-class feature embeddings that are crucial for few-shot learning. Overfitting is a typical issue for few-shot models, the lack of data diversity inhibits proper model inference which weakens the classification accuracy of few-shot learners to novel classes. In this regard, we broke down the pipeline of the few-shot classifier and established that the support, query and task data augmentation collectively alleviates overfitting in networks. With compelling results, we demonstrated that combining a DBT-based low-rank orthogonal regularizer with data augmentation strategies, significantly boosts the performance of a few-shot classifier. We perform our experiments on the miniImagenet, CIFAR-FS and Stanford datasets with performance values of about 5% when compared to state-of-the-art.

Online knowledge distillation (KD) mutually trains a group of student networks from scratch in a peer-teaching manner, eliminating the need for pre-trained teacher models. However, supervision from peers can be noisy, especially in the early stage of training. In this paper, we propose a novel method for online knowledge distillation by temporal-spatial boosting (TSB). The proposed method constructs superior "teachers" with two modules, temporal accumulator and spatial integrator. Specifically, the temporal accumulator leverages the previous outputs of networks during training and produces a representative prediction over all classes. Instead of merely imitating the outputs of other networks as in vanilla online KD, we further propose the so-called spatial integrator that consolidates the knowledge learned by all networks and yields a stronger instructor. The operations of these two modules are simple and straightforward, which can be computed efficiently on the fly during training. The proposed method can improve the efficiency of transferring effective knowledge as well as stabilize the training process. Experimental results on various benchmark datasets and network structures validate the effectiveness of the proposed method over the state-of-the-art.

Continual learning in online scenario aims to learn a sequence of new tasks from data stream using each data only once for training, which is more realistic than in offline mode assuming data from new task are all available. However, this problem is still under-explored for the challenging class-incremental setting in which the model classifies all classes seen so far during inference. Particularly, performance struggles with increased number of tasks or additional classes to learn for each task. In addition, most existing methods require storing original data as exemplars for knowledge replay, which may not be feasible for certain applications with limited memory budget or privacy concerns. In this work, we introduce an effective and memory-efficient method for online continual learning under class-incremental setting through candidates selection from each learned task together with prior incorporation using stored feature embeddings instead of original data as exemplars. Our proposed method implemented for image classification task achieves the best results under different benchmark datasets for online continual learning including CIFAR-10, CIFAR-100 and CORE-50 while requiring much less memory resource compared with existing works.

Low resource Handwritten Text Recognition (HTR) is a hard problem due to the scarce annotated data and the very limited linguistic information (dictionaries and language models). For example, in the case of historical ciphered manuscripts, which are usually written with invented alphabets to hide the message contents. Thus, in this paper we address this problem through a data generation technique based on Bayesian Program Learning (BPL). Contrary to traditional generation approaches, which require a huge amount of annotated images, our method is able to generate human-like handwriting using only one sample of each symbol in the alphabet. After generating symbols, we create synthetic lines to train state-of-the-art HTR architectures in a segmentation free fashion. Quantitative and qualitative analyses were carried out and confirm the effectiveness of the proposed method.