CVPR 2024 Papers

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.

Existing wearable motion capture methods typically demand tight on-body fixation (often using straps) for reliable sensing limiting their application in everyday life. In this paper we introduce Loose Inertial Poser a novel motion capture solution with high wearing comfortableness by integrating four Inertial Measurement Units (IMUs) into a loose-wear jacket. Specifically we address the challenge of scarce loose-wear IMU training data by proposing a Secondary Motion AutoEncoder (SeMo-AE) that learns to model and synthesize the effects of secondary motion between the skin and loose clothing on IMU data. SeMo-AE is leveraged to generate a diverse synthetic dataset of loose-wear IMU data to augment training for the pose estimation network and significantly improve its accuracy. For validation we collected a dataset with various subjects and 2 wearing styles (zipped and unzipped). Experimental results demonstrate that our approach maintains high-quality real-time posture estimation even in loose-wear scenarios.

Neural RGBD SLAM techniques have shown promise in dense Simultaneous Localization And Mapping (SLAM) yet face challenges such as error accumulation during camera tracking resulting in distorted maps. In response we introduce Loopy-SLAM that globally optimizes poses and the dense 3D model. We use frame-to-model tracking using a data-driven point-based submap generation method and trigger loop closures online by performing global place recognition. Robust pose graph optimization is used to rigidly align the local submaps. As our representation is point based map corrections can be performed efficiently without the need to store the entire history of input frames used for mapping as typically required by methods employing a grid based mapping structure. Evaluation on the synthetic Replica and real-world TUM-RGBD and ScanNet datasets demonstrate competitive or superior performance in tracking mapping and rendering accuracy when compared to existing dense neural RGBD SLAM methods. Project page: notchla.github.io/Loopy-SLAM.

Audiovisual representation learning typically relies on the correspondence between sight and sound. However there are often multiple audio tracks that can correspond with a visual scene. Consider for example different conversations on the same crowded street. The effect of such counterfactual pairs on audiovisual representation learning has not been previously explored. To investigate this we use dubbed versions of movies and television shows to augment cross-modal contrastive learning. Our approach learns to represent alternate audio tracks differing only in speech similarly to the same video. Our results from a comprehensive set of experiments investigating different training strategies show this general approach improves performance on a range of downstream auditory and audiovisual tasks without majorly affecting linguistic task performance overall. These findings highlight the importance of considering speech variation when learning scene-level audiovisual correspondences and suggest that dubbed audio can be a useful augmentation technique for training audiovisual models toward more robust performance on diverse downstream tasks.

Automatic anomaly detection based on visual cues holds practical significance in various domains such as manufacturing and product quality assessment. This paper introduces a new conditional anomaly detection problem which involves identifying anomalies in a query image by comparing it to a reference shape. To address this challenge we have created a large dataset BrokenChairs-180K consisting of around 180K images with diverse anomalies geometries and textures paired with 8143 reference 3D shapes. To tackle this task we have proposed a novel transformer-based approach that explicitly learns the correspondence between the query image and reference 3D shape via feature alignment and leverages a customized attention mechanism for anomaly detection. Our approach has been rigorously evaluated through comprehensive experiments serving as a benchmark for future research in this domain.

Anomaly detection (AD) aims to identify defective images and localize their defects (if any). Ideally AD models should be able to detect defects over many image classes; without relying on hard-coded class names that can be uninformative or inconsistent across datasets; learn without anomaly supervision; and be robust to the long-tailed distributions of real-world applications. To address these challenges we formulate the problem of long-tailed AD by introducing several datasets with different levels of class imbalance and metrics for performance evaluation. We then propose a novel method LTAD to detect defects from multiple and long-tailed classes without relying on dataset class names. LTAD combines AD by reconstruction and semantic AD modules. AD by reconstruction is implemented with a transformer-based reconstruction module. Semantic AD is implemented with a binary classifier which relies on learned pseudo class names and a pretrained foundation model. These modules are learned over two phases. Phase 1 learns the pseudo-class names and a variational autoencoder (VAE) for feature synthesis that augments the training data to combat long-tails. Phase 2 then learns the parameters of the reconstruction and classification modules of LTAD. Extensive experiments using the proposed long-tailed datasets show that LTAD substantially outperforms the state-of-the-art methods for most forms of dataset imbalance. The long-tailed dataset split is available at https://zenodo.org/records/10854201

Long-tail class incremental learning (LT-CIL) is designed to perpetually acquire novel knowledge from an imbalanced and perpetually evolving data stream while ensuring the retention of previously acquired knowledge. The existing method only re-balances data distribution and ignores exploring the potential relationship between different samples causing non-robust representations and even severe forgetting in classes with few samples. In this paper we constructed two parallel spaces simultaneously: 1) Sub-prototype space and 2) Reminiscence space to learn robust representations while alleviating forgetfulness. Concretely we advance the concept of the sub-prototype space which amalgamates insights from diverse classes. This integration facilitates the mutual complementarity of varied knowledge thereby augmenting the attainment of more robust representations. Furthermore we introduce the reminiscence space which encapsulates each class distribution aiming to constraint model optimization and mitigate the phenomenon of forgetting. The tandem utilization of the two parallel spaces effectively alleviates the adverse consequences associated with imbalanced data distribution preventing forgetting without needing replay examples. Extensive experiments demonstrate that our method achieves state-of-the-art performance on various benchmarks.

The design of deep network architectures and training methods in computer vision has been well-explored. However in almost all cases the images have been used as provided with little exploration of pre-processing steps beyond normalization and data augmentation. Virtually all images posted on the web or captured by devices are processed for viewing by humans. Is the pipeline used for humans also best for use by computers and deep networks? The human visual system uses logarithmic sensors; differences and sums correspond to ratios and products. Features in log space will be invariant to intensity changes and robust to color balance changes. Log RGB space also reveals structure that is corrupted by typical pre-processing. We explore using linear and log RGB data for training standard backbone architectures on an image classification task using data derived directly from RAW images to guarantee its integrity. We found that networks trained on log RGB data exhibit improved performance on an unmodified test set and invariance to intensity and color balance modifications without additional training or data augmentation. Furthermore we found that the gains from using high quality log data could also be partially or fully realized from data in 8-bit sRGB-JPG format by inverting the sRGB transform and taking the log. These results imply existing databases may benefit from this type of pre-processing. While working with log data we found it was critical to retain the integrity of the log relationships and that networks using log data train best with meta-parameters different than those used for sRGB or linear data. Finally we introduce a new 10-category 10k RAW image data set (RAW10) for image classification and other purposes to enable further the exploration of log RGB as an input format for deep networks in computer vision.

We propose Lodge a network capable of generating extremely long dance sequences conditioned on given music. We design Lodge as a two-stage coarse to fine diffusion architecture and propose the characteristic dance primitives that possess significant expressiveness as intermediate representations between two diffusion models. The first stage is global diffusion which focuses on comprehending the coarse-level music-dance correlation and production characteristic dance primitives. In contrast the second-stage is the local diffusion which parallelly generates detailed motion sequences under the guidance of the dance primitives and choreographic rules. In addition we propose a Foot Refine Block to optimize the contact between the feet and the ground enhancing the physical realism of the motion. Code available at https://li-ronghui.github.io/lodge

We present the Locally Adaptive Morphable Model (LAMM) a highly flexible Auto-Encoder (AE) framework for learning to generate and manipulate 3D meshes. We train our architecture following a simple self-supervised training scheme in which input displacements over a set of sparse control vertices are used to overwrite the encoded geometry in order to transform one training sample into another. During inference our model produces a dense output that adheres locally to the specified sparse geometry while maintaining the overall appearance of the encoded object. This approach results in state-of-the-art performance in both disentangling manipulated geometry and 3D mesh reconstruction. To the best of our knowledge LAMM is the first end-to-end framework that enables direct local control of 3D vertex geometry in a single forward pass. A very efficient computational graph allows our network to train with only a fraction of the memory required by previous methods and run faster during inference generating 12k vertex meshes at >60fps on a single CPU thread. We further leverage local geometry control as a primitive for higher level editing operations and present a set of derivative capabilities such as swapping and sampling object parts. Code and pretrained models can be found at https://github.com/michaeltrs/LAMM.

The task of online mapping is to predict a local map using current sensor observations e.g. from lidar and camera without relying on a pre-built map. State-of-the-art methods are based on supervised learning and are trained predominantly using two datasets: nuScenes and Argoverse 2. However these datasets revisit the same geographic locations across training validation and test sets. Specifically over 80% of nuScenes and 40% of Argoverse 2 validation and test samples are less than 5 m from a training sample. At test time the methods are thus evaluated more on how well they localize within a memorized implicit map built from the training data than on extrapolating to unseen locations. Naturally this data leakage causes inflated performance numbers and we propose geographically disjoint data splits to reveal the true performance in unseen environments. Experimental results show that methods perform considerably worse some dropping more than 45 mAP when trained and evaluated on proper data splits. Additionally a reassessment of prior design choices reveals diverging conclusions from those based on the original split. Notably the impact of lifting methods and the support from auxiliary tasks (e.g. depth supervision) on performance appears less substantial or follows a different trajectory than previously perceived.

Rotation invariance is an important requirement for point shape analysis. To achieve this current state-of-the-art methods attempt to construct the local rotation-invariant representation through learning or defining the local reference frame (LRF). Although efficient these LRF-based methods suffer from perturbation of local geometric relations resulting in suboptimal local rotation invariance. To alleviate this issue we propose a Local-consistent Transformation (LocoTrans) learning strategy. Specifically we first construct the local-consistent reference frame (LCRF) by considering the symmetry of the two axes in LRF. In comparison with previous LRFs our LCRF is able to preserve local geometric relationships better through performing local-consistent transformation. However as the consistency only exists in local regions the relative pose information is still lost in the intermediate layers of the network. We mitigate such a relative pose issue by developing a relative pose recovery (RPR) module. RPR aims to restore the relative pose between adjacent transformed patches. Equipped with LCRF and RPR our LocoTrans is capable of learning local-consistent transformation and preserving local geometry which benefits rotation invariance learning. Competitive performance under arbitrary rotations on both shape classification and part segmentation tasks and ablations can demonstrate the effectiveness of our method. Code will be available publicly at https://github.com/wdttt/LocoTrans.

Referring video object segmentation (RVOS) aims to segment the target instance referred by a given text expression in a video clip. The text expression normally contains sophisticated description of the instance's appearance action and relation with others. It is therefore rather difficult for a RVOS model to capture all these attributes correspondingly in the video; in fact the model often favours more on the action- and relation-related visual attributes of the instance. This can end up with partial or even incorrect mask prediction of the target instance. We tackle this problem by taking a subject-centric short text expression from the original long text expression. The short one retains only the appearance-related information of the target instance so that we can use it to focus the model's attention on the instance's appearance. We let the model make joint predictions using both long and short text expressions; and insert a long-short cross-attention module to interact the joint features and a long-short predictions intersection loss to regulate the joint predictions. Besides the improvement on the linguistic part we also introduce a forward-backward visual consistency loss which utilizes optical flows to warp visual features between the annotated frames and their temporal neighbors for consistency. We build our method on top of two state of the art pipelines. Extensive experiments on A2D-Sentences Refer-YouTube-VOS JHMDB-Sentences and Refer-DAVIS17 show impressive improvements of our method. Code is available here.

Estimating disparities in challenging areas is difficult and limits the performance of stereo matching models. In this paper we exploit local structure information (LSI) to enhance stereo matching. Specifically our LSI comprises a series of key elements including the slant plane (parameterised by disparity gradients) disparity offset details and neighbouring relations. This LSI empowers our method to effectively handle intricate structures including object boundaries and curved surfaces. We bootstrap the LSI from monocular depth and subsequently iteratively refine it to better capture the underlying scene geometry constraints. Building upon the LSI we introduce the Local Structure-Guided Propagation (LSGP) which enhances the disparity initialization optimization and refinement processes. By combining LSGP with a Gated Recurrent Unit (GRU) we present our novel stereo matching method referred to as Local Structure-guided stereo matching (LoS). Remarkably LoS achieves top-ranking results on four widely recognized public benchmark datasets (ETH3D Middlebury KITTI 15 & 12) demonstrating the superior capabilities of our proposed model.

Active Speaker Detection (ASD) aims to identify who is speaking in each frame of a video. Solving ASD involves using audio and visual information in two complementary contexts: long-term intra-speaker context models the temporal dependencies of the same speaker while short-term inter-speaker context models the interactions of speakers in the same scene. Motivated by these observations we propose LoCoNet a simple but effective Long-Short Context Network that leverages Long-term Intra-speaker Modeling (LIM) and Short-term Inter-speaker Modeling (SIM) in an interleaved manner. LIM employs self-attention for long-range temporal dependencies modeling and cross-attention for audio-visual interactions modeling. SIM incorporates convolutional blocks that capture local patterns for short-term inter-speaker context. Experiments show that LoCoNet achieves state-of-the-art performance on multiple datasets with 95.2% (+0.3%) mAP on AVA-ActiveSpeaker 97.2% (+2.7%) mAP on Talkies and 68.4% (+7.7%) mAP on Ego4D. Moreover in challenging cases where multiple speakers are present LoCoNet outperforms previous state-of-the-art methods by 3.0% mAP on AVA-ActiveSpeaker. The code is available at https://github.com/SJTUwxz/LoCoNet_ASD.

The ability to learn from context with novel concepts and deliver appropriate responses are essential in human conversations. Despite current Multimodal Large Language Models (MLLMs) and Large Language Models (LLMs) being trained on mega-scale datasets recognizing unseen images or understanding novel concepts in a training-free manner remains a challenge. In-Context Learning (ICL) explores training-free few-shot learning where models are encouraged to "learn to learn" from limited tasks and generalize to unseen tasks. In this work we propose link-context learning (LCL) which emphasizes "reasoning from cause and effect" to augment the learning capabilities of MLLMs. LCL goes beyond traditional ICL by explicitly strengthening the causal relationship between the support set and the query set. By providing demonstrations with causal links LCL guides the model to discern not only the analogy but also the underlying causal associations between data points which empowers MLLMs to recognize unseen images and understand novel concepts more effectively. To facilitate the evaluation of this novel approach we introduce the ISEKAI dataset comprising exclusively of unseen generated image-label pairs designed for link-context learning. Extensive experiments show that our LCL-MLLM exhibits strong link-context learning capabilities to novel concepts over vanilla MLLMs.