CVPR 2022 Papers

While classic video anomaly detection (VAD) requires labeled normal videos for training, emerging unsupervised VAD (UVAD) aims to discover anomalies directly from fully unlabeled videos. However, existing UVAD methods still rely on shallow models to perform detection or initialization, and they are evidently inferior to classic VAD methods. This paper proposes a full deep neural network (DNN) based solution that can realize highly effective UVAD. First, we, for the first time, point out that deep reconstruction can be surprisingly effective for UVAD, which inspires us to unveil a property named "normality advantage", i.e., normal events will enjoy lower reconstruction loss when DNN learns to reconstruct unlabeled videos. With this property, we propose Localization based Reconstruction (LBR) as a strong UVAD baseline and a solid foundation of our solution. Second, we propose a novel self-paced refinement (SPR) scheme, which is synthesized into LBR to conduct UVAD. Unlike ordinary self-paced learning that injects more samples in an easy-to-hard manner, the proposed SPR scheme gradually drops samples so that suspicious anomalies can be removed from the learning process. In this way, SPR consolidates normality advantage and enables better UVAD in a more proactive way. Finally, we further design a variant solution that explicitly takes the motion cues into account. The solution evidently enhances the UVAD performance, and it sometimes even surpasses the best classic VAD methods. Experiments show that our solution not only significantly outperforms existing UVAD methods by a wide margin (5% to 9% AUROC), but also enables UVAD to catch up with the mainstream performance of classic VAD.

Digital watermarking is widely used for copyright protection. Traditional 3D watermarking approaches or commercial software are typically designed to embed messages into 3D meshes, and later retrieve the messages directly from distorted/undistorted watermarked 3D meshes. However, in many cases, users only have access to rendered 2D images instead of 3D meshes. Unfortunately, retrieving messages from 2D renderings of 3D meshes is still challenging and underexplored. We introduce a novel end-to-end learning framework to solve this problem through: 1) an encoder to covertly embed messages in both mesh geometry and textures; 2) a differentiable renderer to render watermarked 3D objects from different camera angles and under varied lighting conditions; 3) a decoder to recover the messages from 2D rendered images. From our experiments, we show that our model can learn to embed information visually imperceptible to humans, and to retrieve the embedded information from 2D renderings that undergo 3D distortions. In addition, we demonstrate that our method can also work with other renderers, such as ray tracers and real-time renderers with and without fine-tuning.

Both accuracy and efficiency are crucial for image captioning in real-world scenarios. Although Transformer-based models have gained significant improved captioning performance, their computational cost is very high. A feasible way to reduce the time complexity is to exit the prediction early in internal decoding layers without passing the entire model. However, it is not straightforward to devise early exiting into image captioning due to the following issues. On one hand, the representation in shallow layers lacks high-level semantic and sufficient cross-modal fusion information for accurate prediction. On the other hand, the exiting decisions made by internal classifiers are unreliable sometimes. To solve these issues, we propose DeeCap framework for efficient image captioning, which dynamically selects proper-sized decoding layers from a global perspective to exit early. The key to successful early exiting lies in the specially designed imitation learning mechanism, which predicts the deep layer activation with shallow layer features. By deliberately merging the imitation learning into the whole image captioning architecture, the imitated deep layer representation can mitigate the loss brought by the missing of actual deep layers when early exiting is undertaken, resulting in significant reduction in calculation cost with small sacrifice of accuracy. Experiments on the MS COCO and Flickr30k datasets demonstrate the DeeCap can achieve competitive performances with 4 speed-up. Code is available at: https://github.com/feizc/DeeCap.

Decoupling spatiotemporal representation refers to decomposing the spatial and temporal features into dimension-independent factors. Although previous RGB-D-based motion recognition methods have achieved promising performance through the tightly coupled multi-modal spatiotemporal representation, they still suffer from (i) optimization difficulty under small data setting due to the tightly spatiotemporal-entangled modeling; (ii) information redundancy as it usually contains lots of marginal information that is weakly relevant to classification; and (iii) low interaction between multi-modal spatiotemporal information caused by insufficient late fusion. To alleviate these drawbacks, we propose to decouple and recouple spatiotemporal representation for RGB-D-based motion recognition. Specifically, we disentangle the task of learning spatiotemporal representation into 3 sub-tasks: (1) Learning high-quality and dimension independent features through a decoupled spatial and temporal modeling network. (2) Recoupling the decoupled representation to establish stronger space-time dependency. (3) Introducing a Cross-modal Adaptive Posterior Fusion (CAPF) mechanism to capture cross-modal spatiotemporal information from RGB-D data. Seamless combination of these novel designs forms a robust spatiotemporal representation and achieves better performance than state-of-the-art methods on four public motion datasets. Our code is available at https://github.com/damo-cv/MotionRGBD.

Zero-shot semantic segmentation (ZS3) aims to segment the novel categories that have not been seen in the training. Existing works formulate ZS3 as a pixel-level zero-shot classification problem, and transfer semantic knowledge from seen classes to unseen ones with the help of language models pre-trained only with texts. While simple, the pixel-level ZS3 formulation shows the limited capability to integrate vision-language models that are often pre-trained with image-text pairs and currently demonstrate great potential for vision tasks. Inspired by the observation that humans often perform segment-level semantic labeling, we propose to decouple the ZS3 into two sub-tasks: 1) a class-agnostic grouping task to group the pixels into segments. 2) a zero-shot classification task on segments. The former task does not involve category information and can be directly transferred to group pixels for unseen classes. The latter task performs at segment-level and provides a natural way to leverage large-scale vision-language models pre-trained with image-text pairs (e.g. CLIP) for ZS3. Based on the decoupling formulation, we propose a simple and effective zero-shot semantic segmentation model, called ZegFormer, which outperforms the previous methods on ZS3 standard benchmarks by large margins, e.g., 22 points on the PAS-CAL VOC and 3 points on the COCO-Stuff in terms of mIoU for unseen classes. Code will be released at https://github.com/dingjiansw101/ZegFormer.

Weakly supervised learning can help local feature methods to overcome the obstacle of acquiring a large-scale dataset with densely labeled correspondences. However, since weak supervision cannot distinguish the losses caused by the detection and description steps, directly conducting weakly supervised learning within a joint training describe-then-detect pipeline suffers limited performance. In this paper, we propose a decoupled training describe-then-detect pipeline tailored for weakly supervised local feature learning. Within our pipeline, the detection step is decoupled from the description step and postponed until discriminative and robust descriptors are learned. In addition, we introduce a line-to-window search strategy to explicitly use the camera pose information for better descriptor learning. Extensive experiments show that our method, namely PoSFeat (Camera Pose Supervised Feature), outperforms previous fully and weakly supervised methods and achieves state-ofthe-art performance on a wide range of downstream tasks.

This paper probes intrinsic factors behind typical failure cases (e.g spatial inconsistency and boundary confusion) produced by the existing state-of-the-art method in face parsing. To tackle these problems, we propose a novel Decoupled Multi-task Learning with Cyclical Self-Regulation (DML-CSR) for face parsing. Specifically, DML-CSR designs a multi-task model which comprises face parsing, binary edge, and category edge detection. These tasks only share low-level encoder weights without high-level interactions between each other, enabling to decouple auxiliary modules from the whole network at the inference stage. To address spatial inconsistency, we develop a dynamic dual graph convolutional network to capture global contextual information without using any extra pooling operation. To handle boundary confusion in both single and multiple face scenarios, we exploit binary and category edge detection to jointly obtain generic geometric structure and fine-grained semantic clues of human faces. Besides, to prevent noisy labels from degrading model generalization during training, cyclical self-regulation is proposed to self-ensemble several model instances to get a new model and the resulting model then is used to self-distill subsequent models, through alternating iterations. Experiments show that our method achieves the new state-of-the-art performance on the Helen, CelebAMask-HQ, and Lapa datasets. The source code is available at https://github.com/deepinsight/insightface/tree/master/parsing/dml_csr.

State-of-the-art distillation methods are mainly based on distilling deep features from intermediate layers, while the significance of logit distillation is greatly overlooked. To provide a novel viewpoint to study logit distillation, we reformulate the classical KD loss into two parts, i.e., target class knowledge distillation (TCKD) and non-target class knowledge distillation (NCKD). We empirically investigate and prove the effects of the two parts: TCKD transfers knowledge concerning the "difficulty" of training samples, while NCKD is the prominent reason why logit distillation works. More importantly, we reveal that the classical KD loss is a coupled formulation, which (1) suppresses the effectiveness of NCKD and (2) limits the flexibility to balance these two parts. To address these issues, we present Decoupled Knowledge Distillation(DKD), enabling TCKD and NCKD to play their roles more efficiently and flexibly. Compared with complex feature-based methods, our DKD achieves comparable or even better results and has better training efficiency on CIFAR-100, ImageNet, and MS-COCO datasets for image classification and object detection tasks. This paper proves the great potential of logit distillation, and we hope it will be helpful for future research. The code is available at https://github.com/megvii-research/mdistiller.

Neural Radiance Field (NeRF) has gained considerable attention recently for 3D scene reconstruction and novel view synthesis due to its remarkable synthesis quality. However, image blurriness caused by defocus or motion, which often occurs when capturing scenes in the wild, significantly degrades its reconstruction quality. To address this problem, We propose Deblur-NeRF, the first method that can recover a sharp NeRF from blurry input. We adopt an analysis-by-synthesis approach that reconstructs blurry views by simulating the blurring process, thus making NeRF robust to blurry inputs. The core of this simulation is a novel Deformable Sparse Kernel (DSK) module that models spatially-varying blur kernels by deforming a canonical sparse kernel at each spatial location. The ray origin of each kernel point is jointly optimized, inspired by the physical blurring process. This module is parameterized as an MLP that has the ability to be generalized to various blur types. Jointly optimizing the NeRF and the DSK module allows us to restore a sharp NeRF. We demonstrate that our method can be used on both camera motion blur and defocus blur: the two most common types of blur in real scenes. Evaluation results on both synthetic and real-world data show that our method outperforms several baselines. The synthetic and real datasets along with the source code can be find in https://limacv.github.io/deblurnerf/

This work studies the bias issue of pseudo-labeling, a natural phenomenon that widely occurs but often overlooked by prior research. Pseudo-labels are generated when a classifier trained on source data is transferred to unlabeled target data. We observe heavy long-tailed pseudo-labels when a semi-supervised learning model FixMatch predicts labels on the unlabeled set even though the unlabeled data is curated to be balanced. Without intervention, the training model inherits the bias from the pseudo-labels and end up being sub-optimal. To eliminate the model bias, we propose a simple yet effective method DebiasMatch, comprising of an adaptive debiasing module and an adaptive marginal loss. The strength of debiasing and the size of margins can be automatically adjusted by making use of an online updated queue. Benchmarked on ImageNet-1K, DebiasMatch significantly outperforms previous state-of-the-arts by more than 26% and 10.5% on semi-supervised learning (0.2% annotated data) and zero-shot learning tasks respectively.

Transformers have been successfully applied to computer vision due to its powerful modelling capacity with self-attention. However, the good performance of transformers heavily depends on enormous training images. Thus, a data-efficient transformer solution is urgently needed. In this work, we propose an early knowledge distillation framework, which is termed as DearKD, to improvethe data-efficiency required by transformers. Our DearKD is a two-stage framework that first distills the inductive biases from the early intermediate layers of a CNN and then gives the transformer full play by training without distillation. Further, our DearKD can also be applied to the extreme data-free case where no real images are available, where we propose a boundary-preserving intra-divergence loss based on DeepInversion to further close the performance gap against the full-data counterpart. Extensive experiments on ImageNet, partial ImageNet, data-free setting and other downstream tasks prove the superiority of DearKD over its baselines and state-of-the-art methods.

With increasing focus on augmented and virtual reality applications (XR) comes the demand for algorithms that can lift objects from images and videos into representations that are suitable for a wide variety of related 3D tasks. Large-scale deployment of XR devices and applications means that we cannot solely rely on supervised learning, as collecting and annotating data for the unlimited variety of objects in the real world is infeasible. We present a weakly supervised method that is able to decompose a single image of an object into shape (depth and normals), material (albedo, reflectivity and shininess) and global lighting parameters. For training, the method only relies on a rough initial shape estimate of the training objects to bootstrap the learning process. This shape supervision can come for example from a pretrained depth network or - more generically - from a traditional structure-from-motion pipeline. In our experiments, we show that the method can successfully de-render 2D images into a decomposed 3D representation and generalizes to unseen object categories. Since in-the-wild evaluation is difficult due to the lack of ground truth data, we also introduce a photo-realistic synthetic test set that allows for quantitative evaluation. Please find our project page at: https://github.com/Brummi/derender3d

Deep Neural Networks (DNNs) usually have a large number of parameters and consume a huge volume of storage space, which limits the application of DNNs on memory-constrained devices. Network quantization is an appealing way to compress DNNs. However, most of existing quantization methods require the training dataset and a fine-tuning procedure to preserve the quality of a full-precision model. These are unavailable for the confidential scenarios due to personal privacy and security problems. Focusing on this issue, we propose a novel data-free method for network compression called PNMQ, which employs the Parametric Non-uniform Mixed precision Quantization to generate a quantized network. During the compression stage, the optimal parametric non-uniform quantization grid is calculated directly for each layer to minimize the quantization error. User can directly specify the required compression ratio of a network, which is used by the PNMQ algorithm to select bitwidths of layers. This method does not require any model retraining or expensive calculations, which allows efficient implementations for network compression on edge devices. Extensive experiments have been conducted on various computer vision tasks and the results demonstrate that PNMQ achieves better performance than other state-of-the-art methods of network compression.

A typical pipeline for multi-object tracking (MOT) is to use a detector for object localization, and following re-identification (re-ID) for object association. This pipeline is partially motivated by recent progress in both object detection and re-ID, and partially motivated by biases in existing tracking datasets, where most objects tend to have distinguishing appearance and re-ID models are sufficient for establishing associations. In response to such bias, we would like to re-emphasize that methods for multi-object tracking should also work when object appearance is not sufficiently discriminative. To this end, we propose a large-scale dataset for multi-human tracking, where humans have similar appearance, diverse motion and extreme articulation. As the dataset contains mostly group dancing videos, we name it "DanceTrack". We expect DanceTrack to provide a better platform to develop more MOT algorithms that rely less on visual discrimination and depend more on motion analysis. We benchmark several state-of-the-art trackers on our dataset and observe a significant performance drop on DanceTrack when compared against existing benchmarks. The dataset, project code and competition is released at: https://github.com/DanceTrack.

To perform adversarial attacks in the physical world, many studies have proposed adversarial camouflage, a method to hide a target object by applying camouflage patterns on 3D object surfaces. For obtaining optimal physical adversarial camouflage, previous studies have utilized the so-called neural renderer, as it supports differentiability. However, existing neural renderers cannot fully represent various real-world transformations due to a lack of control of scene parameters compared to the legacy photo-realistic renderers. In this paper, we propose the Differentiable Transformation Attack (DTA), a framework for generating a robust physical adversarial pattern on a target object to camouflage it against object detection models with a wide range of transformations. It utilizes our novel Differentiable Transformation Network (DTN), which learns the expected transformation of a rendered object when the texture is changed while preserving the original properties of the target object. Using our attack framework, an adversary can gain both the advantages of the legacy photo-realistic renderers including various physical-world transformations and the benefit of white-box access by offering differentiability. Our experiments show that our camouflaged 3D vehicles can successfully evade state-of-the-art object detection models in the photo-realistic environment (i.e., CARLA on Unreal Engine). Furthermore, our demonstration on a scaled Tesla Model 3 proves the applicability and transferability of our method to the real world.

With the wide applications of deep neural network models in various computer vision tasks, more and more works study the model vulnerability to adversarial examples. For data-free black box attack scenario, existing methods are inspired by the knowledge distillation, and thus usually train a substitute model to learn knowledge from the target model using generated data as input. However, the substitute model always has a static network structure, which limits the attack ability for various target models and tasks. In this paper, we propose a novel dynamic substitute training attack method to encourage substitute model to learn better and faster from the target model. Specifically, a dynamic substitute structure learning strategy is proposed to adaptively generate optimal substitute model structure via a dynamic gate according to different target models and tasks. Moreover, we introduce a task-driven graph-based structure information learning constrain to improve the quality of generated training data, and facilitate the substitute model learning structural relationships from the target model multiple outputs. Extensive experiments have been conducted to verify the efficacy of the proposed attack method, which can achieve better performance compared with the state-of-the-art competitors on several datasets. Project page: https://wxwangiris.github.io/DST

Pedestrian counting is a fundamental tool for understanding pedestrian patterns and crowd flow analysis. Existing works (e.g., image-level pedestrian counting, crossline crowd counting et al.) either only focus on the image-level counting or are constrained to the manual annotation of lines. In this work, we propose to conduct the pedestrian counting from a new perspective - Video Individual Counting (VIC), which counts the total number of individual pedestrians in the given video (a person is only counted once). Instead of relying on the Multiple Object Tracking (MOT) techniques, we propose to solve the problem by decomposing all pedestrians into the initial pedestrians who existed in the first frame and the new pedestrians with separate identities in each following frame. Then, an end-to-end Decomposition and Reasoning Network (DRNet) is designed to predict the initial pedestrian count with the density estimation method and reason the new pedestrian's count of each frame with the differentiable optimal transport. Extensive experiments are conducted on two datasets with congested pedestrians and diverse scenes, demonstrating the effectiveness of our method over baselines with great superiority in counting the individual pedestrians. Code: https://github.com/taohan10200/DRNet.

In this paper, we propose a new approach to train Generative Adversarial Networks (GANs) where we deploy a double-oracle framework using the generator and discriminator oracles. GAN is essentially a two-player zero-sum game between the generator and the discriminator. Training GANs is challenging as a pure Nash equilibrium may not exist and even finding the mixed Nash equilibrium is difficult as GANs have a large-scale strategy space. In DO-GAN, we extend the double oracle framework to GANs. We first generalize the players' strategies as the trained models of generator and discriminator from the best response oracles. We then compute the meta-strategies using a linear program. For scalability of the framework where multiple generators and discriminator best responses are stored in the memory, we propose two solutions: 1) pruning the weakly-dominated players' strategies to keep the oracles from becoming intractable; 2) applying continual learning to retain the previous knowledge of the networks. We apply our framework to established GAN architectures such as vanilla GAN, Deep Convolutional GAN, Spectral Normalization GAN and Stacked GAN. Finally, we conduct experiments on MNIST, CIFAR-10 and CelebA datasets and show that DO-GAN variants have significant improvements in both subjective qualitative evaluation and quantitative metrics, compared with their respective GAN architectures.

Video inpainting remains a challenging problem to fill with plausible and coherent content in unknown areas in video frames despite the prevalence of data-driven methods. Although various transformer-based architectures yield promising result for this task, they still suffer from hallucinating blurry contents and long-term spatial-temporal inconsistency. While noticing the capability of discrete representation for complex reasoning and predictive learning, we propose a novel Discrete Latent Transformer (DLFormer) to reformulate video inpainting tasks into the discrete latent space rather the previous continuous feature space. Specifically, we first learn a unique compact discrete codebook and the corresponding autoencoder to represent the target video. Built upon these representative discrete codes obtained from the entire target video, the subsequent discrete latent transformer is capable to infer proper codes for unknown areas under a self-attention mechanism, and thus produces fine-grained content with long-term spatial-temporal consistency. Moreover, we further explicitly enforce the short-term consistency to relieve temporal visual jitters via a temporal aggregation block among adjacent frames. We conduct comprehensive quantitative and qualitative evaluations to demonstrate that our method significantly outperforms other state-of-the-art approaches in reconstructing visually-plausible and spatial-temporal coherent content with fine-grained details