CVPR 2020 Papers

Video prediction is a pixel-wise dense prediction task to infer future frames based on past frames. Missing appearance details and motion blur are still two major problems for current models, leading to image distortion and temporal inconsistency. We point out the necessity of exploring multi-frequency analysis to deal with the two problems. Inspired by the frequency band decomposition characteristic of Human Vision System (HVS), we propose a video prediction network based on multi-level wavelet analysis to uniformly deal with spatial and temporal information. Specifically, multi-level spatial discrete wavelet transform decomposes each video frame into anisotropic sub-bands with multiple frequencies, helping to enrich structural information and reserve fine details. On the other hand, multilevel temporal discrete wavelet transform which operates on time axis decomposes the frame sequence into sub-band groups of different frequencies to accurately capture multifrequency motions under a fixed frame rate. Extensive experiments on diverse datasets demonstrate that our model shows significant improvements on fidelity and temporal consistency over the state-of-the-art works. Source code and videos are available at https://github.com/Bei-Jin/STMFANet.

Recent work has shown that self-attention can serve as a basic building block for image recognition models. We explore variations of self-attention and assess their effectiveness for image recognition. We consider two forms of self-attention. One is pairwise self-attention, which generalizes standard dot-product attention and is fundamentally a set operator. The other is patchwise self-attention, which is strictly more powerful than convolution. Our pairwise self-attention networks match or outperform their convolutional counterparts, and the patchwise models substantially outperform the convolutional baselines. We also conduct experiments that probe the robustness of learned representations and conclude that self-attention networks may have significant benefits in terms of robustness and generalization.

Data aggregation techniques can significantly improve vision-based policy learning within a training environment, e.g., learning to drive in a specific simulation condition. However, as on-policy data is sequentially sampled and added in an iterative manner, the policy can specialize and overfit to the training conditions. For real-world applications, it is useful for the learned policy to generalize to novel scenarios that differ from the training conditions. To improve policy learning while maintaining robustness when training end-to-end driving policies, we perform an extensive analysis of data aggregation techniques in the CARLA environment. We demonstrate how the majority of them have poor generalization performance, and develop a novel approach with empirically better generalization performance compared to existing techniques. Our two key ideas are (1) to sample critical states from the collected on-policy data based on the utility they provide to the learned policy in terms of driving behavior, and (2) to incorporate a replay buffer which progressively focuses on the high uncertainty regions of the policy's state distribution. We evaluate the proposed approach on the CARLA NoCrash benchmark, focusing on the most challenging driving scenarios with dense pedestrian and vehicle traffic. Our approach improves driving success rate by 16% over state-of-the-art, achieving 87% of the expert performance while also reducing the collision rate by an order of magnitude without the use of any additional modality, auxiliary tasks, architectural modifications or reward from the environment.

Unsupervised domain adaptation has received significant attention in recent years. Most of existing works tackle the closed-set scenario, assuming that the source and target domains share the exactly same categories. In practice, nevertheless, a target domain often contains samples of classes unseen in source domain (i.e., unknown class). The extension of domain adaptation from closed-set to such open-set situation is not trivial since the target samples in unknown class are not expected to align with the source. In this paper, we address this problem by augmenting the state-of-the-art domain adaptation technique, Self-Ensembling, with category-agnostic clusters in target domain. Specifically, we present Self-Ensembling with Category-agnostic Clusters (SE-CC) --- a novel architecture that steers domain adaptation with the additional guidance of category-agnostic clusters that are specific to target domain. These clustering information provides domain-specific visual cues, facilitating the generalization of Self-Ensembling for both closed-set and open-set scenarios. Technically, clustering is firstly performed over all the unlabeled target samples to obtain the category-agnostic clusters, which reveal the underlying data space structure peculiar to target domain. A clustering branch is capitalized on to ensure that the learnt representation preserves such underlying structure by matching the estimated assignment distribution over clusters to the inherent cluster distribution for each target sample. Furthermore, SE-CC enhances the learnt representation with mutual information maximization. Extensive experiments are conducted on Office and VisDA datasets for both open-set and closed-set domain adaptation, and superior results are reported when comparing to the state-of-the-art approaches.

In this paper, we tackle the domain adaptive object detection problem, where the main challenge lies in significant domain gaps between source and target domains. Previous work seeks to plainly align image-level and instance-level shifts to eventually minimize the domain discrepancy. However, they still overlook to match crucial image regions and important instances across domains, which will strongly affect domain shift mitigation. In this work, we propose a simple but effective categorical regularization framework for alleviating this issue. It can be applied as a plug-and-play component on a series of Domain Adaptive Faster R-CNN methods which are prominent for dealing with domain adaptive detection. Specifically, by integrating an image-level multi-label classifier upon the detection backbone, we can obtain the sparse but crucial image regions corresponding to categorical information, thanks to the weakly localization ability of the classification manner. Meanwhile, at the instance level, we leverage the categorical consistency between image-level predictions (by the classifier) and instance-level predictions (by the detection head) as a regularization factor to automatically hunt for the hard aligned instances of target domains. Extensive experiments of various domain shift scenarios show that our method obtains a significant performance gain over original Domain Adaptive Faster R-CNN detectors. Furthermore, qualitative visualization and analyses can demonstrate the ability of our method for attending on the key regions/instances targeting on domain adaptation. Our code is open-source and available at https://github.com/Megvii-Nanjing/CR-DA-DET.

Fully supervised object detection has achieved great success in recent years. However, abundant bounding boxes annotations are needed for training a detector for novel classes. To reduce the human labeling effort, we propose a novel webly supervised object detection (WebSOD) method for novel classes which only requires the web images without further annotations. Our proposed method combines bottom-up and top-down cues for novel class detection. Within our approach, we introduce a bottom-up mechanism based on the well-trained fully supervised object detector (i.e. Faster RCNN) as an object region estimator for web images by recognizing the common objectiveness shared by base and novel classes. With the estimated regions on the web images, we then utilize the top-down attention cues as the guidance for region classification. Furthermore, we propose a residual feature refinement (RFR) block to tackle the domain mismatch between web domain and the target domain. We demonstrate our proposed method on PASCAL VOC dataset with three different novel/base splits. Without any target-domain novel-class images and annotations, our proposed webly supervised object detection model is able to achieve promising performance for novel classes. Moreover, we also conduct transfer learning experiments on large scale ILSVRC 2013 detection dataset and achieve state-of-the-art performance.

Recently, ensemble models have demonstrated empirical capabilities to alleviate the adversarial vulnerability. In this paper, we exploit first-order interactions within ensembles to formalize a reliable and practical defense. We introduce a scenario of interactions that certifiably improves the robustness according to the size of the ensemble, the diversity of the gradient directions, and the balance of the member's contribution to the robustness. We present a joint gradient phase and magnitude regularization (GPMR) as a vigorous approach to impose the desired scenario of interactions among members of the ensemble. Through extensive experiments, including gradient-based and gradient-free evaluations on several datasets and network architectures, we validate the practical effectiveness of the proposed approach compared to the previous methods. Furthermore, we demonstrate that GPMR is orthogonal to other defense strategies developed for single classifiers and their combination can further improve the robustness of ensembles.

Multiview recognition has been well studied in the literature and achieves decent performance in object recognition and retrieval task. However, most previous works rely on supervised learning and some impractical underlying assumptions, such as the availability of all views in training and inference time. In this work, the problem of multiview self-supervised learning (MV-SSL) is investigated, where only image to object association is given. Given this setup, a novel surrogate task for self-supervised learning is proposed by pursuing "object invariant" representation. This is solved by randomly selecting an image feature of an object as object prototype, accompanied with multiview consistency regularization, which results in view invariant stochastic prototype embedding (VISPE). Experiments shows that the recognition and retrieval results using VISPE outperform that of other self-supervised learning methods on seen and unseen data. VISPE can also be applied to semi-supervised scenario and demonstrates robust performance with limited data available. Code is available at https://github.com/chihhuiho/VISPE

This paper presents a method to interpret the success of knowledge distillation by quantifying and analyzing task-relevant and task-irrelevant visual concepts that are encoded in intermediate layers of a deep neural network (DNN). More specifically, three hypotheses are proposed as follows. 1. Knowledge distillation makes the DNN learn more visual concepts than learning from raw data. 2. Knowledge distillation ensures that the DNN is prone to learning various visual concepts simultaneously. Whereas, in the scenario of learning from raw data, the DNN learns visual concepts sequentially. 3. Knowledge distillation yields more stable optimization directions than learning from raw data. Accordingly, we design three types of mathematical metrics to evaluate feature representations of the DNN. In experiments, we diagnosed various DNNs, and above hypotheses were verified.

A new paradigm is proposed for autonomous driving. The new paradigm lies between the end-to-end and pipelined approaches, and is inspired by how humans solve the problem. While it relies on scene understanding, the latter only considers objects that could originate hazard. These are denoted as action inducing, since changes in their state should trigger vehicle actions. They also define a set of explanations for these actions, which should be produced jointly with the latter. An extension of the BDD100K dataset, annotated for a set of 4 actions and 21 explanations, is proposed. A new multi-task formulation of the problem, which optimizes the accuracy of both action commands and explanations, is then introduced. A CNN architecture is finally proposed to solve this problem, by combining reasoning about action inducing objects and global scene context. Experimental results show that the requirement of explanations improves the recognition of action-inducing objects, which in turn leads to better action predictions.

Normalization techniques are important in different advanced neural networks and different tasks. This work investigates a novel dynamic learning-to-normalize (L2N) problem by proposing Exemplar Normalization (EN), which is able to learn different normalization methods for different convolutional layers and image samples of a deep network. EN significantly improves the flexibility of the recently proposed switchable normalization (SN), which solves a static L2N problem by linearly combining several normalizers in each normalization layer (the combination is the same for all samples). Instead of directly employing a multi-layer perceptron (MLP) to learn data-dependent parameters as conditional batch normalization (cBN) did, the internal architecture of EN is carefully designed to stabilize its optimization, leading to many appealing benefits. (1) EN enables different convolutional layers, image samples, categories, benchmarks, and tasks to use different normalization methods, shedding light on analyzing them in a holistic view. (2) EN is effective for various network architectures and tasks. (3) It could replace any normalization layers in a deep network and still produce stable model training. Extensive experiments demonstrate the effectiveness of EN in a wide spectrum of tasks including image recognition, noisy label learning, and semantic segmentation. For example, by replacing BN in the ordinary ResNet50, improvement produced by EN is 300% more than that of SN on both ImageNet and the noisy WebVision dataset. The codes and models will be released.

Event cameras sense intensity changes and have many advantages over conventional cameras. To take advantage of event cameras, some methods have been proposed to reconstruct intensity images from event streams. However, the outputs are still in low resolution (LR), noisy, and unrealistic. The low-quality outputs stem broader applications of event cameras, where high spatial resolution (HR) is needed as well as high temporal resolution, dynamic range, and no motion blur. We consider the problem of reconstructing and super-resolving intensity images from pure events, when no ground truth (GT) HR images and down-sampling kernels are available. To tackle the challenges, we propose a novel end-to-end pipeline that reconstructs LR images from event streams, enhances the image qualities and upsamples the enhanced images, called EventSR. For the absence of real GT images, our method is primarily unsupervised, deploying adversarial learning. To train EventSR, we create an open dataset including both real-world and simulated scenes. The use of both datasets boosts up the network performance, and the network architectures and various loss functions in each phase help improve the image qualities. The whole pipeline is trained in three phases. While each phase is mainly for one of the three tasks, the networks in earlier phases are fine-tuned by respective loss functions in an end-to-end manner. Experimental results show that EventSR generates high-quality SR images from events for both simulated and real-world data.

This paper presents a novel method for labeling real-world neuromorphic camera sensor data by calculating the likelihood of generating an event at each pixel within a short time window, which we refer to as "event probability mask" or EPM. Its applications include (i) objective benchmarking of event denoising performance, (ii) training convolutional neural networks for noise removal called "event denoising convolutional neural network" (EDnCNN), and (iii) estimating internal neuromorphic camera parameters. We provide the first dataset (DVSNOISE20) of real-world labeled neuromorphic camera events for noise removal.

Weakly-supervised object localization (WSOL) has gained popularity over the last years for its promise to train localization models with only image-level labels. Since the seminal WSOL work of class activation mapping (CAM), the field has focused on how to expand the attention regions to cover objects more broadly and localize them better. However, these strategies rely on full localization supervision to validate hyperparameters and for model selection, which is in principle prohibited under the WSOL setup. In this paper, we argue that WSOL task is ill-posed with only image-level labels, and propose a new evaluation protocol where full supervision is limited to only a small held-out set not overlapping with the test set. We observe that, under our protocol, the five most recent WSOL methods have not made a major improvement over the CAM baseline. Moreover, we report that existing WSOL methods have not reached the few-shot learning baseline, where the full-supervision at validation time is used for model training instead. Based on our findings, we discuss some future directions for WSOL.

With the rapid growth of visual content, deep learning to hash is gaining popularity in the image retrieval community recently. Although it greatly facilitates search efficiency, privacy is also at risks when images on the web are retrieved at a large scale and exploited as a rich mine of personal information. An adversary can extract private images by querying similar images from the targeted category for any usable model. Existing methods based on image processing preserve privacy at a sacrifice of perceptual quality. In this paper, we propose a new mechanism based on adversarial examples to "stash" private images in the deep hash space while maintaining perceptual similarity. We first find that a simple approach of hamming distance maximization is not robust against brute-force adversaries. Then we develop a new loss function by maximizing the hamming distance to not only the original category, but also the centers from all the classes, partitioned into clusters of various sizes. The extensive experiment shows that the proposed defense can harden the attacker's efforts by 2-7 orders of magnitude, without significant increase of computational overhead and perceptual degradation. We also demonstrate 30-60% transferability in hash space with a black-box setting. The code is available at: https://github.com/sugarruy/hashstash

We explore the problem of selectively forgetting a particular subset of the data used for training a deep neural network. While the effects of the data to be forgotten can be hidden from the output of the network, insights may still be gleaned by probing deep into its weights. We propose a method for "scrubbing" the weights clean of information about a particular set of training data. The method does not require retraining from scratch, nor access to the data originally used for training. Instead, the weights are modified so that any probing function of the weights is indistinguishable from the same function applied to the weights of a network trained without the data to be forgotten. This condition is a generalized and weaker form of Differential Privacy. Exploiting ideas related to the stability of stochastic gradient descent, we introduce an upper-bound on the amount of information remaining in the weights, which can be estimated efficiently even for deep neural networks.

Low-rank regions capture geometrically meaningful structures in an image which encompass typical local features such as edges, corners and all kinds of regular, symmetric, often repetitive patterns, that are commonly found in man-made environment. While such patterns are challenging current state-of-the-art feature correspondence methods, the recovered homography of a low-rank texture readily provides 3D structure with respect to a 3D plane, without any prior knowledge of the visual information on that plane. However, the automatic and efficient detection of the broad class of low-rank regions is unsolved. Herein, we propose a novel self-supervised low-rank region detection deep network that predicts a low-rank likelihood map from an image. The evaluation of our method on real-world datasets shows not only that it reliably predicts low-rank regions in the image similarly to our baseline method, but thanks to the data augmentations used in the training phase it generalizes well to difficult cases (e.g. day/night lighting, low contrast, underexposure) where the baseline prediction fails.

Object recognition techniques using convolutional neural networks (CNN) have achieved great success. However, state-of-the-art object detection methods still perform poorly on large vocabulary and long-tailed datasets, e.g. LVIS. In this work, we analyze this problem from a novel perspective: each positive sample of one category can be seen as a negative sample for other categories, making the tail categories receive more discouraging gradients. Based on it, we propose a simple but effective loss, named equalization loss, to tackle the problem of long-tailed rare categories by simply ignoring those gradients for rare categories. The equalization loss protects the learning of rare categories from being at a disadvantage during the network parameter updating. Thus the model is capable of learning better discriminative features for objects of rare classes. Without any bells and whistles, our method achieves AP gains of 4.1% and 4.8% for the rare and common categories on the challenging LVIS benchmark, compared to the Mask R-CNN baseline. With the utilization of the effective equalization loss, we finally won the 1st place in the LVIS Challenge 2019. Code has been made available at: https://github.com/tztztztztz/eql.detectron2

We introduce a simple yet effective episode-based training framework for zero-shot learning (ZSL), where the learning system requires to recognize unseen classes given only the corresponding class semantics. During training, the model is trained within a collection of episodes, each of which is designed to simulate a zero-shot classification task. Through training multiple episodes, the model progressively accumulates ensemble experiences on predicting the mimetic unseen classes, which will generalize well on the real unseen classes. Based on this training framework, we propose a novel generative model that synthesizes visual prototypes conditioned on the class semantic prototypes. The proposed model aligns the visual-semantic interactions by formulating both the visual prototype generation and the class semantic inference into an adversarial framework paired with a parameter-economic Multi-modal Cross-Entropy Loss to capture the discriminative information. Extensive experiments on four datasets under both traditional ZSL and generalized ZSL tasks show that our model outperforms the state-of-the-art approaches by large margins.