CVPR 2024 Papers

In this work we present Omni-SMoLA a multimodal architecture that mixes many multi-modal experts efficiently and achieves both high specialist and generalist performance. In contrast to previous models for which we see performance degradation on average when training the models on a wide range of tasks we show that the SMoLA low-rank experts are able to model different skills and task and overall improve the performance of a generalist model. This finding indicates that simple LMM fine-tuning is suboptimal for handling a wide range of tasks and that pairing the act of fine-tuning with specifically-designed architecture changes leads to better performing models.

Nighttime conditions pose a significant challenge to color constancy due to the diversity of lighting conditions and the presence of substantial low-light noise. Existing color constancy methods struggle with nighttime scenes frequently leading to imprecise light color estimations. To tackle nighttime color constancy we propose a novel unsupervised domain adaptation approach that utilizes labeled daytime data to facilitate learning on unlabeled nighttime images. To specifically address the unique lighting conditions of nighttime and ensure the robustness of pseudo labels we propose adaptive channel masking and light uncertainty. By selectively masking channels that are less sensitive to lighting conditions adaptive channel masking directs the model to progressively focus on features less affected by variations in light colors and noise. Additionally our model leverages light uncertainty to provide a pixel-wise uncertainty estimation regarding light color prediction which helps avoid learning from incorrect labels. Our model demonstrates a significant improvement in accuracy achieving 21.5% lower Mean Angular Error (MAE) compared to the state-of-the-art method on our nighttime dataset.

Multimodal learning which integrates data from diverse sensory modes plays a pivotal role in artificial intelligence. However existing multimodal learning methods often struggle with challenges where some modalities appear more dominant than others during multimodal learning resulting in suboptimal performance. To address this challenge we propose MLA (Multimodal Learning with Alternating Unimodal Adaptation). MLA reframes the conventional joint multimodal learning process by transforming it into an alternating unimodal learning process thereby minimizing interference between modalities. Simultaneously it captures cross-modal interactions through a shared head which undergoes continuous optimization across different modalities. This optimization process is controlled by a gradient modification mechanism to prevent the shared head from losing previously acquired information. During the inference phase MLA utilizes a test-time uncertainty-based model fusion mechanism to integrate multimodal information. Extensive experiments are conducted on five diverse datasets encompassing scenarios with complete modalities and scenarios with missing modalities. These experiments demonstrate the superiority of MLA over competing prior approaches. Our code is available at https://github.com/Cecile-hi/Multimodal-Learning-with-Alternating-Unimodal-Adaptation.

Previous multi-task dense prediction methods based on the Mixture of Experts (MoE) have received great performance but they neglect the importance of explicitly modeling the global relations among all tasks. In this paper we present a novel decoder-focused method for multi-task dense prediction called Mixture-of-Low-Rank-Experts (MLoRE). To model the global task relationships MLoRE adds a generic convolution path to the original MoE structure where each task feature can go through this path for explicit parameter sharing. Furthermore to control the parameters and computational cost brought by the increase in the number of experts we take inspiration from LoRA and propose to leverage the low-rank format of a vanilla convolution in the expert network. Since the low-rank experts have fewer parameters and can be dynamically parameterized into the generic convolution the parameters and computational cost do not change much with the increase of experts. Benefiting from this design we increase the number of experts and its reception field to enlarge the representation capacity facilitating multiple dense tasks learning in a unified network. Extensive experiments on the PASCAL-Context and NYUD-v2 benchmarks show that our MLoRE achieves superior performance compared to previous state-of-the-art methods on all metrics. Our code is available at https://github.com/YuqiYang213/MLoRE.

In the academic field the research on human motion prediction tasks mainly focuses on exploiting the observed information to forecast human movements accurately in the near future horizon. However a significant gap appears when it comes to the application field as current models are all trained offline with fixed parameters that are inherently suboptimal to handle the complex yet ever-changing nature of human behaviors. To bridge this gap in this paper we introduce the task of online meta adaptation for human motion prediction based on the insight that finding "smart weights" capable of swift adjustments to suit different motion contexts along the time is a key to improving predictive accuracy. We propose MoML which ingeniously borrows the bilevel optimization spirit of model-agnostic meta-learning to transform previous predictive mistakes into strong inductive biases to guide online adaptation. This is achieved by our MoAdapter blocks that can learn error information by facilitating efficient adaptation via a few gradient steps which fine-tunes our meta-learned "smart" initialization produced by the generic predictor. Considering real-time requirements in practice we further propose Fast-MoML a more efficient variant of MoML that features a closed-form solution instead of conventional gradient update. Experimental results show that our approach can effectively bring many existing offline motion prediction models online and improves their predictive accuracy.

Test-time adaptation (TTA) has emerged as a promising solution to address performance decay due to unforeseen distribution shifts between training and test data. While recent TTA methods excel in adapting to test data variations such adaptability exposes a model to vulnerability against malicious examples an aspect that has received limited attention. Previous studies have uncovered security vulnerabilities within TTA even when a small proportion of the test batch is maliciously manipulated. In response to the emerging threat we propose median batch normalization (MedBN) leveraging the robustness of the median for statistics estimation within the batch normalization layer during test-time inference. Our method is algorithm-agnostic thus allowing seamless integration with existing TTA frameworks. Our experimental results on benchmark datasets including CIFAR10-C CIFAR100-C and ImageNet-C consistently demonstrate that MedBN outperforms existing approaches in maintaining robust performance across different attack scenarios encompassing both instant and cumulative attacks. Through extensive experiments we show that our approach sustains the performance even in the absence of attacks achieving a practical balance between robustness and performance.

Adapting models pre-trained on large-scale datasets to a variety of downstream tasks is a common strategy in deep learning. Consequently parameter-efficient fine-tuning methods have emerged as a promising way to adapt pre-trained models to different tasks while training only a minimal number of parameters. While most of these methods are designed for single-task adaptation parameter-efficient training in Multi-Task Learning (MTL) architectures is still unexplored. In this paper we introduce MTLoRA a novel framework for parameter-efficient training of MTL models. MTLoRA employs Task-Agnostic and Task-Specific Low-Rank Adaptation modules which effectively disentangle the parameter space in MTL fine-tuning thereby enabling the model to adeptly handle both task specialization and interaction within MTL contexts. We applied MTLoRA to hierarchical-transformer-based MTL architectures adapting them to multiple downstream dense prediction tasks. Our extensive experiments on the PASCAL dataset show that MTLoRA achieves higher accuracy on downstream tasks compared to fully fine-tuning the MTL model while reducing the number of trainable parameters by 3.6x. Furthermore MTLoRA establishes a Pareto-optimal trade-off between the number of trainable parameters and the accuracy of the downstream tasks outperforming current state-of-the-art parameter-efficient training methods in both accuracy and efficiency.

In this paper we present a tensor decomposition and low-rank recovery approach (LowRankOcc) for vision-based 3D semantic occupancy prediction. Conventional methods model outdoor scenes with fine-grained 3D grids but the sparsity of non-empty voxels introduces considerable spatial redundancy leading to potential overfitting risks. In contrast our approach leverages the intrinsic low-rank property of 3D occupancy data factorizing voxel representations into low-rank components to efficiently mitigate spatial redundancy without sacrificing performance. Specifically we present the Vertical-Horizontal (VH) decomposition block factorizes 3D tensors into vertical vectors and horizontal matrices. With our "decomposition-encoding-recovery" framework we encode 3D contexts with only 1/2D convolutions and poolings and subsequently recover the encoded compact yet informative context features back to voxel representations. Experimental results demonstrate that LowRankOcc achieves state-of-the-art performances in semantic scene completion on the SemanticKITTI dataset and 3D occupancy prediction on the nuScenes dataset.

Researchers in natural science need reliable methods for quantifying animal behavior. Recently numerous computer vision methods emerged to automate the process. However observing wild species at remote locations remains a challenging task due to difficult lighting conditions and constraints on power supply and data storage. Event cameras offer unique advantages for battery-dependent remote monitoring due to their low power consumption and high dynamic range capabilities. We use this novel sensor to quantify a behavior in Chinstrap penguins called ecstatic display. We formulate the problem as a temporal action detection task determining the start and end times of the behavior. For this purpose we recorded a colony of breeding penguins in Antarctica for several weeks and labeled event data on 16 nests. The developed method consists of a generator of candidate time intervals (proposals) and a classifier of the actions within them. The experiments show that the event cameras' natural response to motion is effective for continuous behavior monitoring and detection reaching a mean average precision (mAP) of 58% (which increases to 63% in good weather conditions). The results also demonstrate the robustness against various lighting conditions contained in the challenging dataset. The low-power capabilities of the event camera allow it to record significantly longer than with a conventional camera. This work pioneers the use of event cameras for remote wildlife observation opening new interdisciplinary opportunities. https:// tub-rip.github.io/ eventpenguins/

For image super-resolution (SR) bridging the gap between the performance on synthetic datasets and real-world degradation scenarios remains a challenge. This work introduces a novel "Low-Res Leads the Way" (LWay) training framework merging Supervised Pre-training with Self-supervised Learning to enhance the adaptability of SR models to real-world images. Our approach utilizes a low-resolution (LR) reconstruction network to extract degradation embeddings from LR images merging them with super-resolved outputs for LR reconstruction. Leveraging unseen LR images for self-supervised learning guides the model to adapt its modeling space to the target domain facilitating fine-tuning of SR models without requiring paired high-resolution (HR) images. The integration of Discrete Wavelet Transform (DWT) further refines the focus on high-frequency details. Extensive evaluations show that our method significantly improves the generalization and detail restoration capabilities of SR models on unseen real-world datasets outperforming existing methods. Our training regime is universally compatible requiring no network architecture modifications making it a practical solution for real-world SR applications.

Parameter-efficient fine-tuning for pre-trained Vision Transformers aims to adeptly tailor a model to downstream tasks by learning a minimal set of new adaptation parameters while preserving the frozen majority of pre-trained parameters. Striking a balance between retaining the generalizable representation capacity of the pre-trained model and acquiring task-specific features poses a key challenge. Currently there is a lack of focus on guiding this delicate trade-off. In this study we approach the problem from the perspective of Singular Value Decomposition (SVD) of pre-trained parameter matrices providing insights into the tuning dynamics of existing methods. Building upon this understanding we propose a Residual-based Low-Rank Rescaling (RLRR) fine-tuning strategy. This strategy not only enhances flexibility in parameter tuning but also ensures that new parameters do not deviate excessively from the pre-trained model through a residual design. Extensive experiments demonstrate that our method achieves competitive performance across various downstream image classification tasks all while maintaining comparable new parameters. We believe this work takes a step forward in offering a unified perspective for interpreting existing methods and serves as motivation for the development of new approaches that move closer to effectively considering the crucial trade-off mentioned above. Our code is available at https://github.com/zstarN70/RLRR.git.

The popularity of large-scale pre-training has promoted the development of medical foundation models. However some studies have shown that although foundation models exhibit strong general feature extraction capabilities their performance on specific tasks is still inferior to task-specific methods. In this paper we explore a new perspective called "Knowledge Decomposition" to improve the performance on specific medical tasks which deconstruct the foundation model into multiple lightweight expert models each dedicated to a particular task with the goal of improving specialization while concurrently mitigating resource expenditure. To accomplish the above objective we design a novel framework named Low-Rank Knowledge Decomposition (LoRKD) which explicitly separates graidents by incorporating low-rank expert modules and the efficient knowledge separation convolution. Extensive experimental results demonstrate that the decomposed models perform well in terms of performance and transferability even surpassing the original foundation models. Source code is available at: https://github.com/MediaBrain-SJTU/LoRKD

This paper focuses on the high computational complexity in Large Language Models (LLMs) a significant challenge in both natural language processing (NLP) and multi-modal tasks. We propose Low-Rank Approximation for Sparse At- tention (LoRA-Sparse) an innovative approach that strate- gically reduces this complexity. LoRA-Sparse introduces low-rank linear projection layers for sparse attention ap- proximation. It utilizes an order-mimic training methodol- ogy which is crucial for efficiently approximating the self- attention mechanism in LLMs. We empirically show that sparse attention not only reduces computational demands but also enhances model performance in both NLP and multi-modal tasks. This surprisingly shows that redundant attention in LLMs might be non-beneficial. We extensively validate LoRA-Sparse through rigorous empirical studies in both (NLP) and multi-modal tasks demonstrating its effec- tiveness and general applicability. Based on LLaMA and LLaVA models our methods can reduce more than half of the self-attention computation with even better performance than full-attention baselines.

The rise of new video modalities like virtual reality or autonomous driving has increased the demand for efficient multi-view video compression methods both in terms of rate-distortion (R-D) performance and in terms of delay and runtime. While most recent stereo video compression approaches have shown promising performance they compress left and right views sequentially leading to poor parallelization and runtime performance. This work presents Low-Latency neural codec for Stereo video Streaming (LLSS) a novel parallel stereo video coding method designed for fast and efficient low-latency stereo video streaming. Instead of using a sequential cross-view motion compensation like existing methods LLSS introduces a bidirectional feature shifting module to directly exploit mutual information among views and encode them effectively with a joint cross-view prior model for entropy coding. Thanks to this design LLSS processes left and right views in parallel minimizing latency; all while substantially improving R-D performance compared to both existing neural and conventional codecs.

Large-scale high-resolution (HR) land-cover mapping is a vital task to survey the Earth's surface and resolve many challenges facing humanity. However it is still a non-trivial task hindered by complex ground details various landforms and the scarcity of accurate training labels over a wide-span geographic area. In this paper we propose an efficient weakly supervised framework (Paraformer) to guide large-scale HR land-cover mapping with easy-access historical land-cover data of low resolution (LR). Specifically existing land-cover mapping approaches reveal the dominance of CNNs in preserving local ground details but still suffer from insufficient global modeling in various landforms. Therefore we design a parallel CNN-Transformer feature extractor in Paraformer consisting of a downsampling-free CNN branch and a Transformer branch to jointly capture local and global contextual information. Besides facing the spatial mismatch of training data a pseudo-label-assisted training (PLAT) module is adopted to reasonably refine LR labels for weakly supervised semantic segmentation of HR images. Experiments on two large-scale datasets demonstrate the superiority of Paraformer over other state-of-the-art methods for automatically updating HR land-cover maps from LR historical labels.

Images suffer from heavy spatial redundancy because pixels in neighboring regions are spatially correlated. Existing approaches strive to overcome this limitation by reducing less meaningful image regions. However current leading methods rely on supervisory signals. They may compel models to preserve content that aligns with labeled categories and discard content belonging to unlabeled categories. This categorical inductive bias makes these methods less effective in real-world scenarios. To address this issue we propose a self-supervised framework for image redundancy reduction called Learning to Rank Patches (LTRP). We observe that image reconstruction of masked image modeling models is sensitive to the removal of visible patches when the masking ratio is high (e.g. 90%). Building upon it we implement LTRP via two steps: inferring the semantic density score of each patch by quantifying variation between reconstructions with and without this patch and learning to rank the patches with the pseudo score. The entire process is self-supervised thus getting out of the dilemma of categorical inductive bias. We design extensive experiments on different datasets and tasks. The results demonstrate that LTRP outperforms both supervised and other self-supervised methods due to the fair assessment of image content.

Generative vision-language models (VLMs) have shown impressive performance in zero-shot vision-language tasks like image captioning and visual question answering.However improving their zero-shot reasoning typically requires second-stage instruction tuning which relies heavily on human-labeled or large language model-generated annotation incurring high labeling costs. To tackle this challenge we introduce Image-Conditioned Caption Correction (ICCC) a novel pre-training task designed to enhance VLMs' zero-shot performance without the need for labeled task-aware data. The ICCC task compels VLMs to rectify mismatches between visual and language concepts thereby enhancing instruction following and text generation conditioned on visual inputs. Leveraging language structure and a lightweight dependency parser we construct data samples of ICCC task from image-text datasets with low labeling and computation costs. Experimental results on BLIP-2 and InstructBLIP demonstrate significant improvements in zero-shot image-text generation-based VL tasks through ICCC instruction tuning.

Diffusion models are just at a tipping point for image super-resolution task. Nevertheless it is not trivial to capitalize on diffusion models for video super-resolution which necessitates not only the preservation of visual appearance from low-resolution to high-resolution videos but also the temporal consistency across video frames. In this paper we propose a novel approach pursuing Spatial Adaptation and Temporal Coherence (SATeCo) for video super-resolution. SATeCo pivots on learning spatial-temporal guidance from low-resolution videos to calibrate both latent-space high-resolution video denoising and pixel-space video reconstruction. Technically SATeCo freezes all the parameters of the pre-trained UNet and VAE and only optimizes two deliberately-designed spatial feature adaptation (SFA) and temporal feature alignment (TFA) modules in the decoder of UNet and VAE. SFA modulates frame features via adaptively estimating affine parameters for each pixel guaranteeing pixel-wise guidance for high-resolution frame synthesis. TFA delves into feature interaction within a 3D local window (tubelet) through self-attention and executes cross-attention between tubelet and its low-resolution counterpart to guide temporal feature alignment. Extensive experiments conducted on the REDS4 and Vid4 datasets demonstrate the effectiveness of our approach.

Most domain adaptation (DA) methods are based on either a convolutional neural networks (CNNs) or a vision transformers (ViTs). They align the distribution differences between domains as encoders without considering their unique characteristics. For instance ViT excels in accuracy due to its superior ability to capture global representations while CNN has an advantage in capturing local representations. This fact has led us to design a hybrid method to fully take advantage of both ViT and CNN called Explicitly Class-specific Boundaries (ECB). ECB learns CNN on ViT to combine their distinct strengths. In particular we leverage ViT's properties to explicitly find class-specific decision boundaries by maximizing the discrepancy between the outputs of the two classifiers to detect target samples far from the source support. In contrast the CNN encoder clusters target features based on the previously defined class-specific boundaries by minimizing the discrepancy between the probabilities of the two classifiers. Finally ViT and CNN mutually exchange knowledge to improve the quality of pseudo labels and reduce the knowledge discrepancies of these models. Compared to conventional DA methods our ECB achieves superior performance which verifies its effectiveness in this hybrid model. The project website can be found https://dotrannhattuong.github.io/ECB/website/.