CVPR 2023 论文列表

Pretraining followed by fine-tuning has proven to be effective in visual recognition tasks. However, fine-tuning all parameters can be computationally expensive, particularly for large-scale models. To mitigate the computational and storage demands, recent research has explored Parameter-Efficient Fine-Tuning (PEFT), which focuses on tuning a minimal number of parameters for efficient adaptation. Existing methods, however, fail to analyze the impact of the additional parameters on the model, resulting in an unclear and suboptimal tuning process. In this paper, we introduce a novel and effective PEFT paradigm, named SNF (Shortcut adaptation via Normalization Flow), which utilizes normalizing flows to adjust the shortcut layers. We highlight that layers without Lipschitz constraints can lead to error propagation when adapting to downstream datasets. Since modifying the over-parameterized residual connections in these layers is expensive, we focus on adjusting the cheap yet crucial shortcuts. Moreover, learning new information with few parameters in PEFT can be challenging, and information loss can result in label information degradation. To address this issue, we propose an information-preserving normalizing flow. Experimental results demonstrate the effectiveness of SNF. Specifically, with only 0.036M parameters, SNF surpasses previous approaches on both the FGVC and VTAB-1k benchmarks using ViT/B-16 as the backbone. The code is available at https://github.com/Wang-Yaoming/SNF

Predicting future action locations is vital for applications like human-robot collaboration. While some computer vision tasks have made progress in predicting human actions, accurately localizing these actions in future frames remains an area with room for improvement. We introduce a new task called spatial action localization in the future (SALF), which aims to predict action locations in both observed and future frames. SALF is challenging because it requires understanding the underlying physics of video observations to predict future action locations accurately. To address SALF, we use the concept of NeuralODE, which models the latent dynamics of sequential data by solving ordinary differential equations (ODE) with neural networks. We propose a novel architecture, AdamsFormer, which extends observed frame features to future time horizons by modeling continuous temporal dynamics through ODE solving. Specifically, we employ the Adams method, a multi-step approach that efficiently uses information from previous steps without discarding it. Our extensive experiments on UCF101-24 and JHMDB-21 datasets demonstrate that our proposed model outperforms existing long-range temporal modeling methods by a significant margin in terms of frame-mAP.

Masked Autoencoders (MAEs) learn generalizable representations for image, text, audio, video, etc., by reconstructing masked input data from tokens of the visible data. Current MAE approaches for videos rely on random patch, tube, or frame based masking strategies to select these tokens. This paper proposes AdaMAE, an adaptive masking strategy for MAEs that is end-to-end trainable. Our adaptive masking strategy samples visible tokens based on the semantic context using an auxiliary sampling network. This network estimates a categorical distribution over spacetime-patch tokens. The tokens that increase the expected reconstruction error are rewarded and selected as visible tokens, motivated by the policy gradient algorithm in reinforcement learning. We show that AdaMAE samples more tokens from the high spatiotemporal information regions, thereby allowing us to mask 95% of tokens, resulting in lower memory requirements and faster pre-training. We conduct ablation studies on the Something-Something v2 (SSv2) dataset to demonstrate the efficacy of our adaptive sampling approach and report state-of-the-art results of 70.0% and 81.7% in top-1 accuracy on SSv2 and Kinetics-400 action classification datasets with a ViT-Base backbone and 800 pre-training epochs. Code and pre-trained models are available at: https://github.com/wgcban/adamae.git

Given the large-scale data and the high annotation cost, pretraining-finetuning becomes a popular paradigm in multiple computer vision tasks. Previous research has covered both the unsupervised pretraining and supervised finetuning in this paradigm, while little attention is paid to exploiting the annotation budget for finetuning. To fill in this gap, we formally define this new active finetuning task focusing on the selection of samples for annotation in the pretraining-finetuning paradigm. We propose a novel method called ActiveFT for active finetuning task to select a subset of data distributing similarly with the entire unlabeled pool and maintaining enough diversity by optimizing a parametric model in the continuous space. We prove that the Earth Mover's distance between the distributions of the selected subset and the entire data pool is also reduced in this process. Extensive experiments show the leading performance and high efficiency of ActiveFT superior to baselines on both image classification and semantic segmentation.

Recently, few-shot action recognition receives increasing attention and achieves remarkable progress. However, previous methods mainly rely on limited unimodal data (e.g., RGB frames) while the multimodal information remains relatively underexplored. In this paper, we propose a novel Active Multimodal Few-shot Action Recognition (AMFAR) framework, which can actively find the reliable modality for each sample based on task-dependent context information to improve few-shot reasoning procedure. In meta-training, we design an Active Sample Selection (ASS) module to organize query samples with large differences in the reliability of modalities into different groups based on modality-specific posterior distributions. In addition, we design an Active Mutual Distillation (AMD) module to capture discriminative task-specific knowledge from the reliable modality to improve the representation learning of unreliable modality by bidirectional knowledge distillation. In meta-test, we adopt Adaptive Multimodal Inference (AMI) module to adaptively fuse the modality-specific posterior distributions with a larger weight on the reliable modality. Extensive experimental results on four public benchmarks demonstrate that our model achieves significant improvements over existing unimodal and multimodal methods.

Transformer-based methods have shown impressive performance in low-level vision tasks, such as image super-resolution. However, we find that these networks can only utilize a limited spatial range of input information through attribution analysis. This implies that the potential of Transformer is still not fully exploited in existing networks. In order to activate more input pixels for better reconstruction, we propose a novel Hybrid Attention Transformer (HAT). It combines both channel attention and window-based self-attention schemes, thus making use of their complementary advantages of being able to utilize global statistics and strong local fitting capability. Moreover, to better aggregate the cross-window information, we introduce an overlapping cross-attention module to enhance the interaction between neighboring window features. In the training stage, we additionally adopt a same-task pre-training strategy to exploit the potential of the model for further improvement. Extensive experiments show the effectiveness of the proposed modules, and we further scale up the model to demonstrate that the performance of this task can be greatly improved. Our overall method significantly outperforms the state-of-the-art methods by more than 1dB. Codes and models are available at https://github.com/XPixelGroup/HAT.

The self-supervised pretraining paradigm has achieved great success in skeleton-based action recognition. However, these methods treat the motion and static parts equally, and lack an adaptive design for different parts, which has a negative impact on the accuracy of action recognition. To realize the adaptive action modeling of both parts, we propose an Actionlet-Dependent Contrastive Learning method (ActCLR). The actionlet, defined as the discriminative subset of the human skeleton, effectively decomposes motion regions for better action modeling. In detail, by contrasting with the static anchor without motion, we extract the motion region of the skeleton data, which serves as the actionlet, in an unsupervised manner. Then, centering on actionlet, a motion-adaptive data transformation method is built. Different data transformations are applied to actionlet and non-actionlet regions to introduce more diversity while maintaining their own characteristics. Meanwhile, we propose a semantic-aware feature pooling method to build feature representations among motion and static regions in a distinguished manner. Extensive experiments on NTU RGB+D and PKUMMD show that the proposed method achieves remarkable action recognition performance. More visualization and quantitative experiments demonstrate the effectiveness of our method.

Test-Time-Training (TTT) is an approach to cope with out-of-distribution (OOD) data by adapting a trained model to distribution shifts occurring at test-time. We propose to perform this adaptation via Activation Matching (ActMAD): We analyze activations of the model and align activation statistics of the OOD test data to those of the training data. In contrast to existing methods, which model the distribution of entire channels in the ultimate layer of the feature extractor, we model the distribution of each feature in multiple layers across the network. This results in a more fine-grained supervision and makes ActMAD attain state of the art performance on CIFAR-100C and Imagenet-C. ActMAD is also architecture- and task-agnostic, which lets us go beyond image classification, and score 15.4% improvement over previous approaches when evaluating a KITTI-trained object detector on KITTI-Fog. Our experiments highlight that ActMAD can be applied to online adaptation in realistic scenarios, requiring little data to attain its full performance.

In contrast to the natural capabilities of humans to learn new tasks in a sequential fashion, neural networks are known to suffer from catastrophic forgetting, where the model's performances on old tasks drop dramatically after being optimized for a new task. Since then, the continual learning (CL) community has proposed several solutions aiming to equip the neural network with the ability to learn the current task (plasticity) while still achieving high accuracy on the previous tasks (stability). Despite remarkable improvements, the plasticity-stability trade-off is still far from being solved, and its underlying mechanism is poorly understood. In this work, we propose Auxiliary Network Continual Learning (ANCL), a novel method that applies an additional auxiliary network which promotes plasticity to the continually learned model which mainly focuses on stability. More concretely, the proposed framework materializes in a regularizer that naturally interpolates between plasticity and stability, surpassing strong baselines on task incremental and class incremental scenarios. Through extensive analyses on ANCL solutions, we identify some essential principles beneath the stability-plasticity trade-off.

Recovering lighting in a scene from a single image is a fundamental problem in computer vision. While a mirror ball light probe can capture omnidirectional lighting, light probes are generally unavailable in everyday images. In this work, we study recovering lighting from accidental light probes (ALPs)---common, shiny objects like Coke cans, which often accidentally appear in daily scenes. We propose a physically-based approach to model ALPs and estimate lighting from their appearances in single images. The main idea is to model the appearance of ALPs by photogrammetrically principled shading and to invert this process via differentiable rendering to recover incidental illumination. We demonstrate that we can put an ALP into a scene to allow high-fidelity lighting estimation. Our model can also recover lighting for existing images that happen to contain an ALP.

The state of the arts in vision-language pretraining (VLP) achieves exemplary performance but suffers from high training costs resulting from slow convergence and long training time, especially on large-scale web datasets. An essential obstacle to training efficiency lies in the entangled prediction rate (percentage of tokens for reconstruction) and corruption rate (percentage of corrupted tokens) in masked language modeling (MLM), that is, a proper corruption rate is achieved at the cost of a large portion of output tokens being excluded from prediction loss. To accelerate the convergence of VLP, we propose a new pretraining task, namely, free language modeling (FLM), that enables a 100% prediction rate with arbitrary corruption rates. FLM successfully frees the prediction rate from the tie-up with the corruption rate while allowing the corruption spans to be customized for each token to be predicted. FLM-trained models are encouraged to learn better and faster given the same GPU time by exploiting bidirectional contexts more flexibly. Extensive experiments show FLM could achieve an impressive 2.5x pretraining time reduction in comparison to the MLM-based methods, while keeping competitive performance on both vision-language understanding and generation tasks.

Dataset Distillation (DD), a newly emerging field, aims at generating much smaller but efficient synthetic training datasets from large ones. Existing DD methods based on gradient matching achieve leading performance; however, they are extremely computationally intensive as they require continuously optimizing a dataset among thousands of randomly initialized models. In this paper, we assume that training the synthetic data with diverse models leads to better generalization performance. Thus we propose two model augmentation techniques, i.e. using early-stage models and parameter perturbation to learn an informative synthetic set with significantly reduced training cost. Extensive experiments demonstrate that our method achieves up to 20x speedup and comparable performance on par with state-of-the-art methods.

Learning-based visual relocalizers exhibit leading pose accuracy, but require hours or days of training. Since training needs to happen on each new scene again, long training times make learning-based relocalization impractical for most applications, despite its promise of high accuracy. In this paper we show how such a system can actually achieve the same accuracy in less than 5 minutes. We start from the obvious: a relocalization network can be split in a scene-agnostic feature backbone, and a scene-specific prediction head. Less obvious: using an MLP prediction head allows us to optimize across thousands of view points simultaneously in each single training iteration. This leads to stable and extremely fast convergence. Furthermore, we substitute effective but slow end-to-end training using a robust pose solver with a curriculum over a reprojection loss. Our approach does not require privileged knowledge, such a depth maps or a 3D model, for speedy training. Overall, our approach is up to 300x faster in mapping than state-of-the-art scene coordinate regression, while keeping accuracy on par. Code is available: https://nianticlabs.github.io/ace

Recently, deep neural networks have been successfully applied for image restoration (IR) (e.g., super-resolution, de-noising, de-blurring). Despite their promising performance, running IR networks requires heavy computation. A large body of work has been devoted to addressing this issue by designing novel neural networks or pruning their parameters. However, the common limitation is that while images are saved in a compressed format before being enhanced by IR, prior work does not consider the impact of compression on the IR quality. In this paper, we present AccelIR, a framework that optimizes image compression considering the end-to-end pipeline of IR tasks. AccelIR encodes an image through IR-aware compression that optimizes compression levels across image blocks within an image according to the impact on the IR quality. Then, it runs a lightweight IR network on the compressed image, effectively reducing IR computation, while maintaining the same IR quality and image size. Our extensive evaluation using seven IR networks shows that AccelIR can reduce the computing overhead of super-resolution, de-nosing, and de-blurring by 49%, 29%, and 32% on average, respectively

We introduce algebraic machine reasoning, a new reasoning framework that is well-suited for abstract reasoning. Effectively, algebraic machine reasoning reduces the difficult process of novel problem-solving to routine algebraic computation. The fundamental algebraic objects of interest are the ideals of some suitably initialized polynomial ring. We shall explain how solving Raven's Progressive Matrices (RPMs) can be realized as computational problems in algebra, which combine various well-known algebraic subroutines that include: Computing the Grobner basis of an ideal, checking for ideal containment, etc. Crucially, the additional algebraic structure satisfied by ideals allows for more operations on ideals beyond set-theoretic operations. Our algebraic machine reasoning framework is not only able to select the correct answer from a given answer set, but also able to generate the correct answer with only the question matrix given. Experiments on the I-RAVEN dataset yield an overall 93.2% accuracy, which significantly outperforms the current state-of-the-art accuracy of 77.0% and exceeds human performance at 84.4% accuracy.

Audiovisual automatic speech recognition (AV-ASR) aims to improve the robustness of a speech recognition system by incorporating visual information. Training fully supervised multimodal models for this task from scratch, however is limited by the need for large labelled audiovisual datasets (in each downstream domain of interest). We present AVFormer, a simple method for augmenting audioonly models with visual information, at the same time performing lightweight domain adaptation. We do this by (i) injecting visual embeddings into a frozen ASR model using lightweight trainable adaptors. We show that these can be trained on a small amount of weakly labelled video data with minimum additional training time and parameters. (ii) We also introduce a simple curriculum scheme during training which we show is crucial to enable the model to jointly process audio and visual information effectively; and finally (iii) we show that our model achieves state of the art zero-shot results on three different AV-ASR benchmarks (How2, VisSpeech and Ego4D), while also crucially preserving decent performance on traditional audio-only speech recognition benchmarks (LibriSpeech). Qualitative results show that our model effectively leverages visual information for robust speech recognition.

In this work, we present a multimodal solution to the problem of 4D face reconstruction from monocular videos. 3D face reconstruction from 2D images is an under-constrained problem due to the ambiguity of depth. State-of-the-art methods try to solve this problem by leveraging visual information from a single image or video, whereas 3D mesh animation approaches rely more on audio. However, in most cases (e.g. AR/VR applications), videos include both visual and speech information. We propose AVFace that incorporates both modalities and accurately reconstructs the 4D facial and lip motion of any speaker, without requiring any 3D ground truth for training. A coarse stage estimates the per-frame parameters of a 3D morphable model, followed by a lip refinement, and then a fine stage recovers facial geometric details. Due to the temporal audio and video information captured by transformer-based modules, our method is robust in cases when either modality is insufficient (e.g. face occlusions). Extensive qualitative and quantitative evaluation demonstrates the superiority of our method over the current state-of-the-art.

Face forgery detection becomes increasingly crucial due to the serious security issues caused by face manipulation techniques. Recent studies in deepfake detection have yielded promising results when the training and testing face forgeries are from the same domain. However, the problem remains challenging when one tries to generalize the detector to forgeries created by unseen methods during training. Observing that face manipulation may alter the relation between different facial action units (AU), we propose the Action Units Relation Learning framework to improve the generality of forgery detection. In specific, it consists of the Action Units Relation Transformer (ART) and the Tampered AU Prediction (TAP). The ART constructs the relation between different AUs with AU-agnostic Branch and AU-specific Branch, which complement each other and work together to exploit forgery clues. In the Tampered AU Prediction, we tamper AU-related regions at the image level and develop challenging pseudo samples at the feature level. The model is then trained to predict the tampered AU regions with the generated location-specific supervision. Experimental results demonstrate that our method can achieve state-of-the-art performance in both the in-dataset and cross-dataset evaluations.

3D object detection techniques commonly follow a pipeline that aggregates predicted object central point features to compute candidate points. However, these candidate points contain only positional information, largely ignoring the object-level shape information. This eventually leads to sub-optimal 3D object detection. In this work, we propose AShapeFormer, a semantics-guided object-level shape encoding module for 3D object detection. This is a plug-n-play module that leverages multi-head attention to encode object shape information. We also propose shape tokens and object-scene positional encoding to ensure that the shape information is fully exploited. Moreover, we introduce a semantic guidance sub-module to sample more foreground points and suppress the influence of background points for a better object shape perception. We demonstrate a straightforward enhancement of multiple existing methods with our AShapeFormer. Through extensive experiments on the popular SUN RGB-D and ScanNetV2 dataset, we show that our enhanced models are able to outperform the baselines by a considerable absolute margin of up to 8.1%. Code will be available at https://github.com/ZechuanLi/AShapeFormer