CVPR 2024 Papers

With the growing size of pre-trained models full fine-tuning and storing all the parameters for various downstream tasks is costly and infeasible. In this paper we propose a new parameter-efficient fine-tuning method Gradient-based Parameter Selection (GPS) demonstrating that only tuning a few selected parameters from the pre-trained model while keeping the remainder of the model frozen can generate similar or better performance compared with the full model fine-tuning method. Different from the existing popular and state-of-the-art parameter-efficient fine-tuning approaches our method does not introduce any additional parameters and computational costs during both the training and inference stages. Another advantage is the model-agnostic and non-destructive property which eliminates the need for any other design specific to a particular model. Compared with the full fine-tuning GPS achieves 3.33% (91.78% vs. 88.45% FGVC) and 9.61% (73.1% vs. 65.57% VTAB) improvement of the accuracy with tuning only 0.36% parameters of the pre-trained model on average over 24 image classification tasks; it also demonstrates a significant improvement of 17% and 16.8% in mDice and mIoU respectively on medical image segmentation task. Moreover GPS achieves state-of-the-art performance compared with existing PEFT methods. The code will be available in https://github.com/FightingFighting/GPS.git.

We propose FlashAvatar a novel and lightweight 3D animatable avatar representation that could reconstruct a digital avatar from a short monocular video sequence in minutes and render high-fidelity photo-realistic images at 300FPS on a consumer-grade GPU. To achieve this we maintain a uniform 3D Gaussian field embedded in the surface of a parametric face model and learn extra spatial offset to model non-surface regions and subtle facial details. While full use of geometric priors can capture high-frequency facial details and preserve exaggerated expressions proper initialization can help reduce the number of Gaussians thus enabling super-fast rendering speed. Extensive experimental results demonstrate that FlashAvatar outperforms existing works regarding visual quality and personalized details and is almost an order of magnitude faster in rendering speed. Project page: https://ustc3dv.github.io/FlashAvatar/

Federated learning (FL) promotes decentralized training while prioritizing data confidentiality. However its application on resource-constrained devices is challenging due to the high demand for computation and memory resources to train deep learning models. Neural network pruning techniques such as dynamic pruning could enhance model efficiency but directly adopting them in FL still poses substantial challenges including post-pruning performance degradation high activation memory usage etc. To address these challenges we propose FedMef a novel and memory-efficient federated dynamic pruning framework. FedMef comprises two key components. First we introduce the budget-aware extrusion that maintains pruning efficiency while preserving post-pruning performance by salvaging crucial information from parameters marked for pruning within a given budget. Second we propose scaled activation pruning to effectively reduce activation memory footprints which is particularly beneficial for deploying FL to memory-limited devices. Extensive experiments demonstrate the effectiveness of our proposed FedMef. In particular it achieves a significant reduction of 28.5% in memory footprint compared to state-of-the-art methods while obtaining superior accuracy.

We study supervised action segmentation whose goal is to predict framewise action labels of a video. To capture temporal dependencies over long horizons prior works either improve framewise features with transformer or refine framewise predictions with learned action features. However they are computationally costly and ignore that frame and action features contain complimentary information which can be leveraged to enhance both features and improve temporal modeling. Therefore we propose an efficient Frame-Action Cross-attention Temporal modeling (FACT) framework that performs temporal modeling with frame and action features in parallel and leverage this parallelism to achieve iterative bidirectional information transfer between the features and refine them. FACT network contains (i) a frame branch to learn frame-level information with convolutions and frame features (ii) an action branch to learn action-level dependencies with transformers and action tokens and (iii) cross-attentions to allow communication between the two branches. We also propose a new matching loss to ensure each action token uniquely encodes an action segment thus better captures its semantics. Thanks to our architecture we can also leverage textual transcripts of videos to help action segmentation. We evaluate FACT on four video datasets (two egocentric and two third-person) for action segmentation with and without transcripts showing that it significantly improves the state-of-the-art accuracy while enjoys lower computational cost (3 times faster) than existing transformer-based methods

Self-supervised denoising has attracted widespread attention due to its ability to train without clean images. However noise in real-world scenarios is often spatially correlated which causes many self-supervised algorithms that assume pixel-wise independent noise to perform poorly. Recent works have attempted to break noise correlation with downsampling or neighborhood masking. However denoising on downsampled subgraphs can lead to aliasing effects and loss of details due to a lower sampling rate. Furthermore the neighborhood masking methods either come with high computational complexity or do not consider local spatial preservation during inference. Through the analysis of existing methods we point out that the key to obtaining high-quality and texture-rich results in real-world self-supervised denoising tasks is to train at the original input resolution structure and use asymmetric operations during training and inference. Based on this we propose Asymmetric Tunable Blind-Spot Network (AT-BSN) where the blind-spot size can be freely adjusted thus better balancing noise correlation suppression and image local spatial destruction during training and inference. In addition we regard the pre-trained AT-BSN as a meta-teacher network capable of generating various teacher networks by sampling different blind-spots. We propose a blind-spot based multi-teacher distillation strategy to distill a lightweight network significantly improving performance. Experimental results on multiple datasets prove that our method achieves state-of-the-art and is superior to other self-supervised algorithms in terms of computational overhead and visual effects.

Dataset distillation has emerged as a promising approach in deep learning enabling efficient training with small synthetic datasets derived from larger real ones. Particularly distribution matching-based distillation methods attract attention thanks to its effectiveness and low computational cost. However these methods face two primary limitations: the dispersed feature distribution within the same class in synthetic datasets reducing class discrimination and an exclusive focus on mean feature consistency lacking precision and comprehensiveness. To address these challenges we introduce two novel constraints: a class centralization constraint and a covariance matching constraint. The class centralization constraint aims to enhance class discrimination by more closely clustering samples within classes. The covariance matching constraint seeks to achieve more accurate feature distribution matching between real and synthetic datasets through local feature covariance matrices particularly beneficial when sample sizes are much smaller than the number of features. Experiments demonstrate notable improvements with these constraints yielding performance boosts of up to 6.6% on CIFAR10 2.9% on SVHN 2.5% on CIFAR100 and 2.5% on TinyImageNet compared to the state-of-the-art relevant methods. In addition our method maintains robust performance in cross-architecture settings with a maximum performance drop of 1.7% on four architectures. Code is available at https://github.com/VincenDen/IID.

Weakly-supervised action segmentation is a task of learning to partition a long video into several action segments where training videos are only accompanied by transcripts (ordered list of actions). Most of existing methods need to infer pseudo segmentation for training by serial alignment between all frames and the transcript which is time-consuming and hard to be parallelized while training. In this work we aim to escape from this inefficient alignment with massive but redundant frames and instead to directly localize a few action transitions for pseudo segmentation generation where a transition refers to the change from an action segment to its next adjacent one in the transcript. As the true transitions are submerged in noisy boundaries due to intra-segment visual variation we propose a novel Action-Transition-Aware Boundary Alignment (ATBA) framework to efficiently and effectively filter out noisy boundaries and detect transitions. In addition to boost the semantic learning in the case that noise is inevitably present in the pseudo segmentation we also introduce video-level losses to utilize the trusted video-level supervision. Extensive experiments show the effectiveness of our approach on both performance and training speed.

The paradigm of pre-training and fine-tuning has laid the foundation for deploying deep learning models. However most fine-tuning methods are designed to meet a specific resource budget. Recently considering diverse deployment scenarios with various resource budgets SN-Net is introduced to quickly obtain numerous new networks (stitches) from the pre-trained models (anchors) in a model family via model stitching. Although promising SN-Net confronts new challenges when adapting it to new target domains including huge memory and storage requirements and a long and sub-optimal multistage adaptation process. In this work we present a novel framework Efficient Stitchable Task Adaptation (ESTA) to efficiently produce a palette of fine-tuned models that adhere to diverse resource constraints. Specifically we first tailor parameter-efficient fine-tuning to share low-rank updates among the stitches while maintaining independent bias terms. In this way we largely reduce fine-tuning memory burdens and mitigate the interference among stitches that arises in task adaptation. Furthermore we streamline a simple yet effective one-stage deployment pipeline which estimates the important stitches to deploy with training-time gradient statistics. By assigning higher sampling probabilities to important stitches we also get a boosted Pareto frontier. Extensive experiments on 25 downstream visual recognition tasks demonstrate that our ESTA is capable of generating stitches with smooth accuracy-efficiency trade-offs and surpasses the direct SN-Net adaptation by remarkable margins with significantly lower training time and fewer trainable parameters. Furthermore we demonstrate the flexibility and scalability of our ESTA framework by stitching LLMs from LLaMA family obtaining chatbot stitches of assorted sizes.

We revisit certain problems of pose estimation based on 3D--2D correspondences between features which may be points or lines. Specifically we address the two previously-studied minimal problems of estimating camera extrinsics from p \in \ 1 2 \ point--point correspondences and l=3-p line--line correspondences. To the best of our knowledge all of the previously-known practical solutions to these problems required computing the roots of degree \ge 4 (univariate) polynomials when p=2 or degree \ge 8 polynomials when p=1. We describe and implement two elementary solutions which reduce the degrees of the needed polynomials from 4 to 2 and from 8 to 4 respectively. We show experimentally that the resulting solvers are numerically stable and fast: when compared to the previous state-of-the art we may obtain nearly an order of magnitude speedup. The code is available at https://github.com/petrhruby97/efficient_absolute

Scene recovery the restoration of images degraded by adverse weather conditions presents significant challenges for existing methods. Physical models constrained by their inherent assumptions often fail when these assumptions are not met; Deep learning models are powerful they are limited by the diversity of their training datasets leading to poor generalization and high computational demands. To address these limitations we propose the Luminous Flux Prior (LFP) to recover degraded images under diverse adverse weather without learning. Luminous flux a physical measure that reflects image brightness has a rate of change that demonstrates a significant correlation with transmission. Consequently we leverage this rate of change in luminous flux as prior knowledge to estimate transmission which in turn assists in image recovery. This approach reduces dependency on physical parameters and enhances adaptability to various weather. Experimental validation under diverse conditions such as sandstorms underwater environments and haze attests to the robustness of LFP in restoring clear images. With a time complexity of \mathcal O (N\log N) LFP enables real-time recovery making it a suitable for devices with limited computational resources.

The recent success in revealing scene details from sparse 3D point clouds obtained via structure-from-motion has raised significant privacy concerns in visual localization. One prominent approach for mitigating this issue is to lift 3D points to 3D lines thereby reducing the effectiveness of the scene inversion attacks but this comes at the cost of increased algorithmic complexity for camera localization due to weaker geometric constraints induced by line clouds. To overcome this limitation we propose a new lifting approach called "ray cloud" whereby each lifted 3D line intersects at one of two predefined locations depicting omnidirectional rays from two cameras. This yields two benefits i) camera localization can now be cast as relative pose estimation between the query image and the calibrated rig of two perspective cameras which can be efficiently solved using a variant of the 5-point algorithm and ii) the ray cloud introduces erroneous estimations for the density-based inversion attack degrading the quality of scene recovery. Moreover we explore possible modifications of the inversion attack to better recover scenes from the ray clouds and propose a ray sampling technique to reduce the effectiveness of the modified attack. Experimental results on two public datasets show real-time localization speed as well as enhanced privacy-preserving capability over the state-of-the-art without overly sacrificing the localization accuracy.

Multi-task learning for dense prediction has emerged as a pivotal area in computer vision enabling simultaneous processing of diverse yet interrelated pixel-wise prediction tasks. However the substantial computational demands of state-of-the-art (SoTA) models often limit their widespread deployment. This paper addresses this challenge by introducing network binarization to compress resource-intensive multi-task dense predictors. Specifically our goal is to significantly accelerate multi-task dense prediction models via Binary Neural Networks (BNNs) while maintaining and even improving model performance at the same time. To reach this goal we propose a Binary Multi-task Dense Predictor Bi-MTDP and several variants of \bimtdp in which a multi-task dense predictor is constructed via specified binarized modules. Our systematical analysis of this predictor reveals that performance drop from binarization is primarily caused by severe information degradation. To address this issue we introduce a deep information bottleneck layer that enforces representations for downstream tasks satisfying Gaussian distribution in forward propagation. Moreover we introduce a knowledge distillation mechanism to correct the direction of information flow in backward propagation. Intriguingly one variant of Bi-MTDP outperforms full-precision (FP) multi-task dense prediction SoTAs ARTC (CNN-based) and InvPT (ViT-based). This result indicates that Bi-MTDP is not merely a naive trade-off between performance and efficiency but is rather a benefit of the redundant information flow thanks to the multi-task architecture.

Coarse-to-fine schemes are widely used in traditional single-image motion deblur; however in the context of deep learning existing multi-scale algorithms not only require the use of complex modules for feature fusion of low-scale RGB images and deep semantics but also manually generate low-resolution pairs of images that do not have sufficient confidence. In this work we propose a multi-scale network based on single-input and multiple-outputs(SIMO) for motion deblurring. This simplifies the complexity of algorithms based on a coarse-to-fine scheme. To alleviate restoration defects impacting detail information brought about by using a multi-scale architecture we combine the characteristics of real-world blurring trajectories with a learnable wavelet transform module to focus on the directional continuity and frequency features of the step-by-step transitions between blurred images to sharp images. In conclusion we propose a multi-scale network with a learnable discrete wavelet transform (MLWNet) which exhibits state-of-the-art performance on multiple real-world deblurred datasets in terms of both subjective and objective quality as well as computational efficiency.

With the escalating complexity and investment cost of training deep neural networks safeguarding them from unauthorized usage and intellectual property theft has become imperative. Especially the rampant misuse of prediction APIs to replicate models without access to the original data or architecture poses grave security threats. Diverse defense strategies have emerged to address these vulnerabilities yet these defenses either incur heavy inference overheads or assume idealized attack scenarios. To address these challenges we revisit the utilization of noise transition matrix as an efficient perturbation technique which injects noise into predicted posteriors in a linear manner and integrates seamlessly into existing systems with minimal overhead for model stealing defense. Provably with such perturbed posteriors the attacker's cloning process degrades into learning from noisy data. Toward optimizing the noise transition matrix we proposed a novel bi-level optimization training framework which performs fidelity on the victim model while the surrogate model adversarially. Comprehensive experimental results demonstrate that our method effectively thwarts model stealing attacks and achieves minimal utility trade-offs outperforming existing state-of-the-art defenses.

In this paper we explore the problem of event-based meshflow estimation a novel task that involves predicting a spatially smooth sparse motion field from event cameras. To start we generate a large-scale High-Resolution Event Meshflow (HREM) dataset which showcases its superiority by encompassing the merits of high resolution at 1280x720 handling dynamic objects and complex motion patterns and offering both optical flow and meshflow labels. These aspects have not been fully explored in previous works. Besides we propose Efficient Event-based MeshFlow (EEMFlow) network a lightweight model featuring a specially crafted encoder-decoder architecture to facilitate swift and accurate meshflow estimation. Furthermore we upgrade EEMFlow network to support dense event optical flow in which a Confidence-induced Detail Completion (CDC) module is proposed to preserve sharp motion boundaries. We conduct comprehensive experiments to show the exceptional performance and runtime efficiency (39x faster) of our EEMFlow model compared to recent state-of-the-art flow methods. Our code is available at https://github.com/boomluo02/EEMFlow.

Hyperparameter Optimization and Neural Architecture Search are powerful in attaining state-of-the-art machine learning models with Bayesian Optimization (BO) standing out as a mainstream method. Extending BO into the multi-fidelity setting has been an emerging research topic in this field but faces the challenge of determining an appropriate fidelity for each hyperparameter configuration to fit the surrogate model. To tackle the challenge we propose a multi-fidelity BO method named FastBO which excels in adaptively deciding the fidelity for each configuration and providing strong performance while ensuring efficient resource usage. These advantages are achieved through our proposed techniques based on the concepts of efficient point and saturation point for each configuration which can be obtained from the empirical learning curve of the configuration estimated from early observations. Extensive experiments demonstrate FastBO's superior anytime performance and efficiency in identifying high-quality configurations and architectures. We also show that our method provides a way to extend any single-fidelity method to the multi-fidelity setting highlighting the wide applicability of our approach.

3D head avatars built with neural implicit volumetric representations have achieved unprecedented levels of photorealism. However the computational cost of these methods remains a significant barrier to their widespread adoption particularly in real-time applications such as virtual reality and teleconferencing. While attempts have been made to develop fast neural rendering approaches for static scenes these methods cannot be simply employed to support realistic facial expressions such as in the case of a dynamic facial performance. To address these challenges we propose a novel fast 3D neural implicit head avatar model that achieves real-time rendering while maintaining fine-grained controllability and high rendering quality. Our key idea lies in the introduction of local hash table blendshapes which are learned and attached to the vertices of an underlying face parametric model. These per-vertex hash-tables are linearly merged with weights predicted via a CNN resulting in expression dependent embeddings. Our novel representation enables efficient density and color predictions using a lightweight MLP which is further accelerated by a hierarchical nearest neighbor search method. Extensive experiments show that our approach runs in real-time while achieving comparable rendering quality to state-of-the-arts and decent results on challenging expressions.

Collaborative perception enhances perception performance by enabling autonomous vehicles to exchange complementary information. Despite its potential to revolutionize the mobile industry challenges in various environments such as communication bandwidth limitations localization errors and information aggregation inefficiencies hinder its implementation in practical applications. In this work we propose ERMVP a communication-Efficient and collaboration-Robust Multi-Vehicle Perception method in challenging environments. Specifically ERMVP has three distinct strengths: i) It utilizes the hierarchical feature sampling strategy to abstract a representative set of feature vectors using less communication overhead for efficient communication; ii) It employs the sparse consensus features to execute precise spatial location calibrations effectively mitigating the implications of vehicle localization errors; iii) A pioneering feature fusion and interaction paradigm is introduced to integrate holistic spatial semantics among different vehicles and data sources. To thoroughly validate our method we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that the proposed ERMVP is significantly superior to the state-of-the-art collaborative perception methods.