CVPR 2022 Papers

Recent transformer-based solutions have been introduced to estimate 3D human pose from 2D keypoint sequence by considering body joints among all frames globally to learn spatio-temporal correlation. We observe that the motions of different joints differ significantly. However, the previous methods cannot efficiently model the solid inter-frame correspondence of each joint, leading to insufficient learning of spatial-temporal correlation. We propose MixSTE (Mixed Spatio-Temporal Encoder), which has a temporal transformer block to separately model the temporal motion of each joint and a spatial transformer block to learn inter-joint spatial correlation. These two blocks are utilized alternately to obtain better spatio-temporal feature encoding. In addition, the network output is extended from the central frame to entire frames of the input video, thereby improving the coherence between the input and output sequences. Extensive experiments are conducted on three benchmarks (Human3.6M, MPI-INF-3DHP, and HumanEva). The results show that our model outperforms the state-of-the-art approach by 10.9% P-MPJPE and 7.6% MPJPE. The code is available at https://github.com/JinluZhang1126/MixSTE.

While local-window self-attention performs notably in vision tasks, it suffers from limited receptive field and weak modeling capability issues. This is mainly because it performs self-attention within non-overlapped windows and shares weights on the channel dimension. We propose MixFormer to find a solution. First, we combine local-window self-attention with depth-wise convolution in a parallel design, modeling cross-window connections to enlarge the receptive fields. Second, we propose bi-directional interactions across branches to provide complementary clues in the channel and spatial dimensions. These two designs are integrated to achieve efficient feature mixing among windows and dimensions. Our MixFormer provides competitive results on image classification with EfficientNet and shows better results than RegNet and Swin Transformer. Performance in downstream tasks outperforms its alternatives by significant margins with less computational costs in 5 dense prediction tasks on MS COCO, ADE20k, and LVIS. Code is available at https://github.com/PaddlePaddle/PaddleClas.

Tracking often uses a multi-stage pipeline of feature extraction, target information integration, and bounding box estimation. To simplify this pipeline and unify the process of feature extraction and target information integration, we present a compact tracking framework, termed as MixFormer, built upon transformers. Our core design is to utilize the flexibility of attention operations, and propose a Mixed Attention Module (MAM) for simultaneous feature extraction and target information integration. This synchronous modeling scheme allows to extract target-specific discriminative features and perform extensive communication between target and search area. Based on MAM, we build our MixFormer tracking framework simply by stacking multiple MAMs with progressive patch embedding and placing a localization head on top. In addition, to handle multiple target templates during online tracking, we devise an asymmetric attention scheme in MAM to reduce computational cost, and propose an effective score prediction module to select high-quality templates. Our MixFormer sets a new state-of-the-art performance on five tracking benchmarks, including LaSOT, TrackingNet, VOT2020, GOT-10k, and UAV123. In particular, our MixFormer-L achieves NP score of 79.9% on LaSOT, 88.9% on TrackingNet and EAO of 0.555 on VOT2020. We also perform in-depth ablation studies to demonstrate the effectiveness of simultaneous feature extraction and information integration. Code and trained models are publicly available at https://github.com/MCG-NJU/MixFormer.

We present a method for simultaneously localizing multiple sound sources within a visual scene. This task requires a model to both group a sound mixture into individual sources, and to associate them with a visual signal. Our method jointly solves both tasks at once, using a formulation inspired by the contrastive random walk of Jabri et al. We create a graph in which images and separated sounds each correspond to nodes, and train a random walker to transition between nodes from different modalities with high return probability. The transition probabilities for this walk are determined by an audio-visual similarity metric that is learned by our model. We show through experiments with musical instruments and human speech that our model can successfully localize multiple sounds, outperforming other self-supervised methods.

Though neural radiance fields ("NeRF") have demonstrated impressive view synthesis results on objects and small bounded regions of space, they struggle on "unbounded" scenes, where the camera may point in any direction and content may exist at any distance. In this setting, existing NeRF-like models often produce blurry or low-resolution renderings (due to the unbalanced detail and scale of nearby and distant objects), are slow to train, and may exhibit artifacts due to the inherent ambiguity of the task of reconstructing a large scene from a small set of images. We present an extension of mip-NeRF (a NeRF variant that addresses sampling and aliasing) that uses a non-linear scene parameterization, online distillation, and a novel distortion-based regularizer to overcome the challenges presented by unbounded scenes. Our model, which we dub "mip-NeRF 360" as we target scenes in which the camera rotates 360 degrees around a point, reduces mean-squared error by 57% compared to mip-NeRF, and is able to produce realistic synthesized views and detailed depth maps for highly intricate, unbounded real-world scenes.

In this work, we study the cross-view information fusion problem in the task of self-supervised 3D hand pose estimation from the depth image. Previous methods usually adopt a hand-crafted rule to generate pseudo labels from multi-view estimations in order to supervise the network training in each view. However, these methods ignore the rich semantic information in each view and ignore the complex dependencies between different regions of different views. To solve these problems, we propose a cross-view fusion network to fully exploit and adaptively aggregate multi-view information. We encode diverse semantic information in each view into multiple compact nodes. Then, we introduce the graph convolution to model the complex dependencies between nodes and perform cross-view information interaction. Based on the cross-view fusion network, we propose a strong self-supervised framework for 3D hand pose and hand mesh estimation. Furthermore, we propose a pseudo multi-view training strategy to extend our framework to a more general scenario in which only single-view training data is used. Results on NYU dataset demonstrate that our method outperforms the previous self-supervised methods by 17.5% and 30.3% in multi-view and single-view scenarios. Meanwhile, our framework achieves comparable results to several strongly supervised methods.

Vision Transformer (ViT) models have recently drawn much attention in computer vision due to their high model capability. However, ViT models suffer from huge number of parameters, restricting their applicability on devices with limited computation. To alleviate this problem, we propose MiniViT, a new compression framework, which achieves parameter reduction in vision transformers while retaining the same performance. The central idea of MiniViT is to multiplex the weights of consecutive transformer blocks. More specifically, we make the weights shared across layers, while imposing a transformation on the weights to increase diversity. Weight distillation over self-attention is also applied to transfer knowledge from large-scale ViT models to weight-multiplexed compact models. Comprehensive experiments demonstrate the efficacy of MiniViT, showing that it can reduce the size of the pre-trained Swin-B transformer by 48%, while achieving an increase of 1.0% in top-1 accuracy on ImageNet. Moreover, using a single-layer parameters, MiniViT is able to compress DeiT-B by 9.7 times from 86M to 9M parameters, without seriously compromising the performance. Finally, we verify the transferability of MiniViT by reporting its performance on downstream benchmarks. Code and models are available at here.

Class Incremental Learning (CIL) aims at learning a classifier in a phase-by-phase manner, in which only data of a subset of the classes are provided at each phase. Previous works mainly focus on mitigating forgetting in phases after the initial one. However, we find that improving CIL at its initial phase is also a promising direction. Specifically, we experimentally show that directly encouraging CIL Learner at the initial phase to output similar representations as the model jointly trained on all classes can greatly boost the CIL performance. Motivated by this, we study the difference between a naively-trained initial-phase model and the oracle model. Specifically, since one major difference between these two models is the number of training classes, we investigate how such difference affects the model representations. We find that, with fewer training classes, the data representations of each class lie in a long and narrow region; with more training classes, the representations of each class scatter more uniformly. Inspired by this observation, we propose Class-wise Decorrelation (CwD) that effectively regularizes representations of each class to scatter more uniformly, thus mimicking the model jointly trained with all classes (i.e., the oracle model). Our CwD is simple to implement and easy to plug into existing methods. Extensive experiments on various benchmark datasets show that CwD consistently and significantly improves the performance of existing state-of-the-art methods by around 1% to 3%. Code: https://github.com/Yujun-Shi/CwD.

In the era of deep learning, human pose estimation from multiple cameras with unknown calibration has received little attention to date. We show how to train a neural model to perform this task with high precision and minimal latency overhead. The proposed model takes into account joint location uncertainty due to occlusion from multiple views, and requires only 2D keypoint data for training. Our method outperforms both classical bundle adjustment and weakly-supervised monocular 3D baselines on the well-established Human3.6M dataset, as well as the more challenging in-the-wild Ski-Pose PTZ dataset.

Transformers have shown great potential in computer vision tasks. A common belief is their attention-based token mixer module contributes most to their competence. However, recent works show the attention-based module in transformers can be replaced by spatial MLPs and the resulted models still perform quite well. Based on this observation, we hypothesize that the general architecture of the transformers, instead of the specific token mixer module, is more essential to the model's performance. To verify this, we deliberately replace the attention module in transformers with an embarrassingly simple spatial pooling operator to conduct only basic token mixing. Surprisingly, we observe that the derived model, termed as PoolFormer, achieves competitive performance on multiple computer vision tasks. For example, on ImageNet-1K, PoolFormer achieves 82.1% top-1 accuracy, surpassing well-tuned vision transformer/MLP-like baselines DeiT-B/ResMLP-B24 by 0.3%/1.1% accuracy with 35%/52% fewer parameters and 49%/61% fewer MACs. The effectiveness of PoolFormer verifies our hypothesis and urges us to initiate the concept of "MetaFormer", a general architecture abstracted from transformers without specifying the token mixer. Based on the extensive experiments, we argue that MetaFormer is the key player in achieving superior results for recent transformer and MLP-like models on vision tasks. This work calls for more future research dedicated to improving MetaFormer instead of focusing on the token mixer modules. Additionally, our proposed PoolFormer could serve as a starting baseline for future MetaFormer architecture design.

In this paper, we tackle the problem of few-shot class incremental learning (FSCIL). FSCIL aims to incrementally learn new classes with only a few samples in each class. Most existing methods only consider the incremental steps at test time. The learning objective of these methods is often hand-engineered and is not directly tied to the objective (i.e. incrementally learning new classes) during testing. Those methods are sub-optimal due to the misalignment between the training objectives and what the methods are expected to do during evaluation. In this work, we proposed a bi-level optimization based on meta-learning to directly optimize the network to learn how to incrementally learn in the setting of FSCIL. Concretely, we propose to sample sequences of incremental tasks from base classes for training to simulate the evaluation protocol. For each task, the model is learned using a meta-objective such that it is capable to perform fast adaptation without forgetting. Furthermore, we propose a bi-directional guided modulation, which is learned to automatically modulate the activations to reduce catastrophic forgetting. Extensive experimental results demonstrate that the proposed method outperforms the baseline and achieves the state-of-the-art results on CIFAR100, MiniImageNet, and CUB200 datasets.

Continual learning is a longstanding research topic due to its crucial role in tackling continually arriving tasks. Up to now, the study of continual learning in computer vision is mainly restricted to convolutional neural networks (CNNs). However, recently there is a tendency that the newly emerging vision transformers (ViTs) are gradually dominating the field of computer vision, which leaves CNN-based continual learning lagging behind as they can suffer from severe performance degradation if straightforwardly applied to ViTs. In this paper, we study ViT-backed continual learning to strive for higher performance riding on recent advances of ViTs. Inspired by mask-based continual learning methods in CNNs, where a mask is learned per task to adapt the pre-trained ViT to the new task, we propose MEta-ATtention (MEAT), i.e., attention to self-attention, to adapt a pre-trained ViT to new tasks without sacrificing performance on already learned tasks. Unlike prior mask-based methods like Piggyback, where all parameters are associated with corresponding masks, MEAT leverages the characteristics of ViTs and only masks a portion of its parameters. It renders MEAT more efficient and effective with less overhead and higher accuracy. Extensive experiments demonstrate that MEAT exhibits significant superiority to its state-of-the-art CNN counterparts, with 4.0 6.0% absolute boosts in accuracy. Our code has been released at https://github.com/zju-vipa/MEAT-TIL.

Domain Generalizable (DG) person ReID is a challenging task which trains a model on source domains yet generalizes well on target domains. Existing methods use source domains to learn domain-invariant features, and assume those features are also irrelevant with target domains. However, they do not consider the target domain information which is unavailable in the training phrase of DG. To address this issue, we propose a novel Meta Distribution Alignment (MDA) method to enable them to share similar distribution in a test-time-training fashion. Specifically, since high-dimensional features are difficult to constrain with a known simple distribution, we first introduce an intermediate latent space constrained to a known prior distribution. The source domain data is mapped to this latent space and then reconstructed back. A meta-learning strategy is introduced to facilitate generalization and support fast adaption. To reduce their discrepancy, we further propose a test-time adaptive updating strategy based on the latent space which efficiently adapts model to unseen domains with a few samples. Extensive experimental results show that our model outperforms the state-of-the-art methods by up to 5.1% R-1 on average on the large-scale and 4.7% R-1 on the single-source domain generalization ReID benchmark.

In single domain generalization, models trained with data from only one domain are required to perform well on many unseen domains. In this paper, we propose a new model, termed meta convolutional neural network, to solve the single domain generalization problem in image recognition. The key idea is to decompose the convolutional features of images into meta features. Acting as "visual words", meta features are defined as universal and basic visual elements for image representations (like words for documents in language). Taking meta features as reference, we propose compositional operations to eliminate irrelevant features of local convolutional features by an addressing process and then to reformulate the convolutional feature maps as a composition of related meta features. In this way, images are universally coded without biased information from the unseen domain, which can be processed by following modules trained in the source domain. The compositional operations adopt a regression analysis technique to learn the meta features in an online batch learning manner. Extensive experiments on multiple benchmark datasets verify the superiority of the proposed model in improving single domain generalization ability.

The existing pose estimation approaches often require a large number of annotated images to attain good estimation performance, which are laborious to acquire. To reduce the human efforts on pose annotations, we propose a novel Meta Agent Teaming Active Learning (MATAL) framework to actively select and label informative images for effective learning. Our MATAL formulates the image selection procedure as a Markov Decision Process and learns an optimal sampling policy that directly maximizes the performance of the pose estimator. Our framework consists of a novel state-action representation as well as a multi-agent team to enable batch sampling in the active learning procedure. The framework could be effectively optimized via Meta-Optimization to accelerate the adaptation to the gradually expanded labeled data during deployment. Finally, we show experimental results on both human hand and body pose estimation benchmark datasets and demonstrate that our method significantly outperforms all baselines continuously under the same amount of annotation budget. Moreover, to obtain similar pose estimation accuracy, our MATAL framework can save around 40% labeling efforts on average compared to state-of-the-art active learning frameworks.

A large proportion of videos captured today are first per-son videos shot from wearable cameras. Similar to other computer vision tasks, Deep Neural Networks (DNNs) are the workhorse for most state-of-the-art (SOTA) egocentric vision techniques. On the other hand DNNs are known to be susceptible to Adversarial Attacks (AAs) which add im-perceptible noise to the input. Both black-box, as well as white-box attacks on image as well as video analysis tasks have been shown. We observe that most AA techniques basically add intensity perturbation to an image. Even for videos, the same process is essentially repeated for each frame independently. We note that the definition of imperceptibility used for images may not be applicable for videos, where a small intensity change happening randomly in two consecutive frames may still be perceptible. In this paper we make a key novel suggestion to use perturbation in optical flow to carry out AAs on a video analysis system. Such perturbation is especially useful for egocentric videos, because there is a lot of shake in the egocentric videos anyways, and adding a little more, keeps it highly imperceptible. In general, our idea can be seen as adding structured, para-metric noise as the adversarial perturbation. Our implementation of the idea by adding 3D rotations to the frames reveal that using our technique, one can mount a black-box AA on an egocentric activity detection system in one-third of the queries compared to the SOTA AA technique.

In this paper, we propose a simple yet effective video super-resolution method that aims at generating high-fidelity high-resolution (HR) videos from low-resolution (LR) ones. Previous methods predominantly leverage temporal neighbor frames to assist the super-resolution of the current frame. Those methods achieve limited performance as they suffer from the challenges in spatial frame alignment and the lack of useful information from similar LR neighbor frames. In contrast, we devise a cross-frame non-local attention mechanism that allows video super-resolution without frame alignment, leading to being more robust to large motions in the video. In addition, to acquire general video prior information beyond neighbor frames, and to compensate for the information loss caused by large motions, we design a novel memory-augmented attention module to memorize general video details during the super-resolution training. We have thoroughly evaluated our work on various challenging datasets. Compared to other recent video super-resolution approaches, our method not only achieves significant performance gains on large motion videos but also shows better generalization. Our source code and the new Parkour benchmark dataset will be released.

Pan-sharpening aims to obtain high-resolution multispectral (MS) images for remote sensing systems and deep learning-based methods have achieved remarkable success. However, most existing methods are designed in a black-box principle, lacking sufficient interpretability. Additionally, they ignore the different characteristics of each band of MS images and directly concatenate them with panchromatic (PAN) images, leading to severe copy artifacts. To address the above issues, we propose an interpretable deep neural network, namely Memory-augmented Deep Conditional Unfolding Network with two specified core designs. Firstly, considering the degradation process, it formulates the Pan-sharpening problem as the minimization of a variational model with denoising-based prior and non-local auto-regression prior which is capable of searching the similarities between long-range patches, benefiting the texture enhancement. A novel iteration algorithm with built-in CNNs is exploited for transparent model design. Secondly, to fully explore the potentials of different bands of MS images, the PAN image is combined with each band of MS images, selectively providing the high-frequency details and alleviating the copy artifacts. Extensive experimental results validate the superiority of the proposed algorithm against other state-of-the-art methods.

We use neural radiance fields (NeRFs) to build interactive 3D environments from large-scale visual captures spanning buildings or even multiple city blocks collected primarily from drones. In contrast to single object scenes (on which NeRFs are traditionally evaluated), our scale poses multiple challenges including (1) the need to model thousands of images with varying lighting conditions, each of which capture only a small subset of the scene, (2) prohibitively large model capacities that make it infeasible to train on a single GPU, and (3) significant challenges for fast rendering that would enable interactive fly-throughs. To address these challenges, we begin by analyzing visibility statistics for large-scale scenes, motivating a sparse network structure where parameters are specialized to different regions of the scene. We introduce a simple geometric clustering algorithm for data parallelism that partitions training images (or rather pixels) into different NeRF submodules that can be trained in parallel. We evaluate our approach on existing datasets (Quad 6k and UrbanScene3D) as well as against our own drone footage, improving training speed by 3x and PSNR by 12%. We also evaluate recent NeRF fast renderers on top of Mega-NeRF and introduce a novel method that exploits temporal coherence. Our technique achieves a 40x speedup over conventional NeRF rendering while remaining within 0.8 db in PSNR quality, exceeding the fidelity of existing fast renderers.