David Marr's seminal theory of human perception stipulates that visual processing is a multi-stage process, prioritizing the derivation of boundary and surface properties before forming semantic object representations. In contrast, contrastive representation learning frameworks typically bypass this explicit multi-stage approach, defining their objective as the direct learning of a semantic representation space for objects. While effective in general contexts, this approach sacrifices the inductive biases of vision, leading to slower convergence speed and learning shortcut resulting in texture bias. In this work, we demonstrate that leveraging Marr's multi-stage theory--by first constructing boundary and surface-level representations using perceptual constructs from early visual processing stages and subsequently training for object semantics--leads to 2x faster convergence on ResNet18, improved final representations on semantic segmentation, depth estimation, and object recognition, and enhanced robustness and out-of-distribution capability. Together, we propose a pretraining stage before the general contrastive representation pretraining to further enhance the final representation quality and reduce the overall convergence time via inductive bias from human vision systems.
论文检索
输入标题、作者或关键词，从 2,871 篇学术成果中精准定位
Training large-scale neural networks in vision, and multimodal domains demands substantial memory resources, primarily due to the storage of optimizer states. While LoRA, a popular parameter-efficient method, reduces memory usage, it often suffers from suboptimal performance due to the constraints of low-rank updates. Low-rank gradient projection methods (e.g., GaLore, Flora) reduce optimizer memory by projecting gradients and moment estimates into low-rank spaces via singular value decomposition or random projection. However, they fail to account for inter-projection correlation, causing performance degradation, and their projection strategies often incur high computational costs. In this paper, we present COAP (Correlation-Aware Gradient Projection), a memory-efficient method that minimizes computational overhead while maintaining training performance. Evaluated across various vision, language, and multimodal tasks, COAP outperforms existing methods in both training speed and model performance. For LLaMA-1B, it reduces optimizer memory by 61% with only 2% additional time cost, achieving the same PPL as AdamW. With 8-bit quantization, COAP cuts optimizer memory by 81% and achieves 4x speedup over GaLore for LLaVA-v1.5-7B fine-tuning, while delivering higher accuracy.
Deep neural networks (DNNs) often underperform in real-world, dynamic settings where data distributions change over time. Domain Incremental Learning (DIL) offers a solution by enabling continuous model adaptation, with Parameter-Isolation DIL (PIDIL) emerging as a promising paradigm to reduce knowledge conflicts. However, existing PIDIL methods struggle with parameter selection accuracy, especially as the number of domains and corresponding classes grows. To address this, we propose SOYO, a lightweight framework that improves domain selection in PIDIL. SOYO introduces a Gaussian Mixture Compressor (GMC) and Domain Feature Resampler (DFR) to store and balance prior domain data efficiently, while a Multi-level Domain Feature Fusion Network (MDFN) enhances domain feature extraction. Our framework supports multiple Parameter-Efficient Fine-Tuning (PEFT) methods and is validated across tasks such as image classification, object detection, and speech enhancement. Experimental results on six benchmarks demonstrate SOYO's consistent superiority over existing baselines, showcasing its robustness and adaptability in complex, evolving environments.
Prompt2Perturb (P2P): Text-Guided Diffusion-Based Adversarial Attack on Breast Ultrasound Images
PDF ↗Deep neural networks (DNNs) offer significant promise for improving breast cancer diagnosis in medical imaging. However, these models are highly susceptible to adversarial attacks--small, imperceptible changes that can mislead classifiers--raising critical concerns about their reliability and security. Traditional attack methods typically either require substantial extra data for malicious model pre-training, or involve a fixed norm perturbation budget, which does not align with human perception of these alterations. In medical imaging, however, this is often unfeasible due to the limited availability of datasets. Building on recent advancements in learnable prompts, we propose Prompt2Perturb (P2P), a novel language-guided semantic attack method capable of generating meaningful perturbations driven by text instructions. During the prompt learning phase, our approach leverages learnable prompts within the text encoder to create subtle, yet impactful, perturbations that remain imperceptible while guiding the model towards targeted outcomes.In contrast to current prompt learning-based approaches, our P2P stands out by directly updating text embeddings, avoiding the need for retraining diffusion models or using large pre-trained models which is typically infeasible in medical domain. Further, we leverage the finding that optimizing only the early reverse diffusion steps boosts efficiency while ensuring that the generated adversarial examples incorporate subtle low-frequency noise, thus preserving ultrasound image quality without introducing noticeable artifacts. We show that our method outperforms state-of-the-art attack techniques across three breast ultrasound datasets. Moreover, the generated images are both more natural in appearance and more effective compared to existing adversarial attacks.
From Alexnet to Transformers: Measuring the Non-linearity of Deep Neural Networks with Affine Optimal Transport
PDF ↗In the last decade, we have witnessed the introduction of several novel deep neural network (DNN) architectures exhibiting ever-increasing performance across diverse tasks. Explaining the upward trend of their performance, however, remains difficult as different DNN architectures of comparable depth and width -- common factors associated with their expressive power -- may exhibit a drastically different performance even when trained on the same dataset. In this paper, we introduce the concept of the non-linearity signature of DNN, the first theoretically sound solution for approximately measuring the non-linearity of deep neural networks. Built upon a score derived from closed-form optimal transport mappings, this signature provides a better understanding of the inner workings of a wide range of DNN architectures and learning paradigms, with a particular emphasis on the computer vision task. We provide extensive experimental results that highlight the practical usefulness of the proposed non-linearity signature and its potential for long-reaching implications.
Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis
PDF ↗Diffusion-based text-to-image (T2I) models have recently excelled in high-quality image generation, particularly in a training-free manner, enabling cost-effective adaptability and generalization across diverse tasks. However, while the existing methods have been continuously focusing on several challenges such as "missing objects" and "mismatched attributes," another critical issue of "mislocated objects" remains where generated spatial positions fail to align with text prompts. Surprisingly, ensuring such seemingly basic functionality remains challenging in popular T2I models due to the inherent difficulty of imposing explicit spatial guidance via text forms. To address this, we propose STORM (Spatial Transport Optimization by Repositioning Attention Map), a novel training-free approach for spatially coherent T2I synthesis. STORM employs Spatial Transport Optimization (STO), rooted in optimal transport theory, to dynamically adjust object attention maps for precise spatial adherence, supported by a Spatial Transport (ST) Cost function that enhances spatial understanding. Our analysis shows that integrating spatial awareness is most effective in the early denoising stages, while later phases refine details. Extensive experiments demonstrate that STORM surpasses existing methods, effectively mitigating mislocated objects while improving missing and mismatched attributes, setting a new benchmark for spatial alignment in T2I synthesis.
Spatio-Temporal Scene Graphs (STSGs) provide a concise and expressive representation of dynamic scenes by modeling objects and their evolving relationships over time. However, real-world visual relationships often exhibit a long-tailed distribution, causing existing methods for tasks like Video Scene Graph Generation (VidSGG) and Scene Graph Anticipation (SGA) to produce biased scene graphs. To this end, we propose ImparTail, a novel training framework that leverages loss masking and curriculum learning to mitigate bias in the generation and anticipation of spatio-temporal scene graphs. Unlike prior methods that add extra architectural components to learn unbiased estimators, we propose an impartial training objective that reduces the dominance of head classes during learning and focuses on underrepresented tail relationships. Our curriculum-driven mask generation strategy further empowers the model to adaptively adjust its bias mitigation strategy over time, enabling more balanced and robust estimations. To thoroughly assess performance under various distribution shifts, we also introduce two new tasks--Robust Spatio-Temporal Scene Graph Generation and Robust Scene Graph Anticipation--offering a challenging benchmark for evaluating the resilience of STSG models. Extensive experiments on the Action Genome dataset demonstrate the superior unbiased performance and robustness of our method compared to existing baselines.
Out-of-distribution (OOD) detection is crucial for ensuring the robustness of machine learning models by identifying samples that deviate from the training distribution. While traditional OOD detection has predominantly focused on single-modality inputs, such as images, recent advancements in multimodal models have shown the potential of utilizing multiple modalities (e.g., video, optical flow, audio) to improve detection performance. However, existing approaches often neglect intra-class variability within in-distribution (ID) data, assuming that samples of the same class are perfectly cohesive and consistent. This assumption can lead to performance degradation, especially when prediction discrepancies are indiscriminately amplified across all samples. To address this issue, we propose Dynamic Prototype Updating (DPU), a novel plug-and-play framework for multimodal OOD detection that accounts for intra-class variations. Our method dynamically updates class center representations for each class by measuring the variance of similar samples within each batch, enabling tailored adjustments. This approach allows us to intensify prediction discrepancies based on the updated class centers, thereby enhancing the model's robustness and generalization across different modalities. Extensive experiments on two tasks, five datasets, and nine base OOD algorithms demonstrate that DPU significantly improves OOD detection performances, setting a new state-of-the-art in multimodal OOD detection, including improvements up to 80% in Far-OOD detection.To improve accessibility and reproducibility, our code is released anonymously at https://anonymous.4open.science/r/CVPR-9177.
Finding semantic correspondences between images is a challenging problem in computer vision, particularly under significant viewpoint changes. Previous methods rely on semantic features from pre-trained 2D models like Stable Diffusion and DINOv2, which often struggle to extract viewpoint-invariant features. To overcome this, we propose a novel approach that integrates geometric and semantic reasoning. Unlike prior methods relying on heuristic geometric enhancements, our framework fine-tunes DUSt3R on synthetic cross-instance data to reconstruct distinct objects in an aligned 3D space. By learning to deform these objects into similar shapes using semantic supervision, we enable efficient KNN-based geometric matching, followed by sparse semantic matching within local KNN candidates. While trained on synthetic data, our method generalizes effectively to real-world images, achieving up to 7.4-point improvements in zero-shot settings on the rigid-body subset of Spair-71K and up to 19.6-point gains under extreme viewpoint variations. Additionally, it accelerates runtime by up to 40 times, demonstrating both its robustness to viewpoint changes and its efficiency for practical applications.
SemAlign3D: Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations
PDF ↗Semantic correspondence made tremendous progress through the recent advancements of large vision models (LVM). While these LVMs have been shown to reliably capture local semantics, the same can currently not be said for capturing global geometric relationships between semantic object regions. This problem leads to unreliable performance for semantic correspondence between images with extreme view variation. In this work, we aim to leverage monocular depth estimates to capture these geometric relationships for more robust and data-efficient semantic correspondence. First, we introduce a simple but effective method to build 3D object-class representations from monocular depth estimates and LVM features using a sparsely annotated image correspondence dataset. Second, we formulate an alignment energy that can be minimized using gradient descent to obtain an alignment between the 3D object-class representation and the object-class instance in the input RGB-image. Our method achieves state-of-the-art matching accuracy in multiple categories on the challenging SPair-71k dataset, increasing the PCK@0.1 score by more than 10 points on three categories and overall by 3.3 points from 85.6% to 88.9%. Additional resources and code are available at dub.sh/semalign3d.
Generative replay has gained significant attention in class-incremental learning; however, its application to Class Incremental Object Detection (CIOD) remains limited due to the challenges in generating complex images with precise spatial arrangements. In this study, motivated by the observation that the forgetting of prior knowledge is predominantly present in the classification sub-task as opposed to the localization sub-task, we revisit the generative replay method for class incremental object detection. Our method utilize a standard Stable Diffusion model to generate image-level replay data for all old and new tasks. Accordingly, the old detector and a stage-wise detector are conducted on the synthetic images respectively to determine the bounding box positions through pseudo-labeling. Furthermore, we propose to use a Similarity-based Cross Sampling mechanism to select more valuable confusing data between old and new tasks to more effectively mitigate catastrophic forgetting and reduce the false alarm rate for the new task. Finally, all synthetic and real data are integrated for current-stage detector training, where the images generated for previous tasks are highly beneficial in minimizing the forgetting of existing knowledge, while those synthesized for the new task can help bridge the domain gap between real and synthetic images. We conducted extensive experiments on PASCAL VOC 2007 and MS COCO benchmark datasets in multiple settings to showcase the efficacy of our proposed approach, which achieves state-of-the-art results. The code are available at https://github.com/qiangzai-lv/RGR-IOD.
DSV-LFS: Unifying LLM-Driven Semantic Cues with Visual Features for Robust Few-Shot Segmentation
PDF ↗Few-shot semantic segmentation (FSS) aims to enable models to segment novel/unseen object classes using only a limited number of labeled examples. However, current FSS methods frequently struggle with generalization due to incomplete and biased feature representations, especially when support images do not capture the full appearance variability of the target class. To improve the FSS pipeline, we propose a novel framework that utilizes large language models (LLMs) to adapt general class semantic information to the query image. Furthermore, the framework employs dense pixel-wise matching to identify similarities between query and support images, resulting in enhanced FSS performance. Inspired by reasoning-based segmentation frameworks, our method, named DSV-LFS, introduces an additional token into the LLM vocabulary, allowing a multimodal LLM to generate a "semantic prompt" from class descriptions. In parallel, a dense matching module identifies visual similarities between the query and support images, generating a "visual prompt". These prompts are then jointly employed to guide the prompt-based decoder for accurate segmentation of the query image. Comprehensive experiments on the benchmark datasets Pascal-5i and COCO-20i demonstrate that our framework achieves state-of-the-art performance-by a significant margin-demonstrating superior generalization to novel classes and robustness across diverse scenarios.
3D affordance segmentation aims to link human instructions to touchable regions of 3D objects for embodied manipulations. Existing efforts typically adhere to single-object, single-affordance paradigms, where each affordance type or explicit instruction strictly corresponds to a specific affordance region and are unable to handle long-horizon tasks. Such a paradigm cannot actively reason about complex user intentions that often imply sequential affordances. In this paper, we introduce the Sequential 3D Affordance Reasoning task, which extends the traditional paradigm by reasoning from cumbersome user intentions and then decomposing them into a series of segmentation maps. Toward this, we construct the first instruction-based affordance segmentation benchmark that includes reasoning over both single and sequential affordances, comprising 180K instruction-point cloud pairs. Based on the benchmark, we propose our model, SeqAfford, to unlock the 3D multi-modal large language model with additional affordance segmentation abilities, which ensures reasoning with world knowledge and fine-grained affordance grounding in a cohesive framework. We further introduce a multi-granular language-point integration module to endow 3D dense prediction. Extensive experimental evaluations show that our model excels over well-established methods and exhibits open-world generalization with sequential reasoning abilities.
Model quantization is an effective way to compress deep neural networks and accelerate the inference time on edge devices. Existing quantization methods usually require original data for calibration during the compressing process, which may be inaccessible due to privacy issues. A common way is to generate calibration data to mimic the origin data. However, the generators in these methods have the mode collapse problem, making them unable to synthesize diverse data. To solve this problem, we leverage the information from the full-precision model and enhance both inter-class and intra-class diversity for generating better calibration data, by devising a multi-layer features mixer and normalization flow based attention. Besides, novel regulation losses are proposed to make the generator produce diverse data with more patterns from the perspective of activated feature values and for the quantized model to learn better clip ranges adaptive to our diverse calibration data. Extensive experiments show that our method achieves state-of-the-art quantization results for both Transformer and CNN architectures. In addition, we visualize the generated data to verify that our strategies can effectively handle the mode collapse issue. Our codes are available at https://anonymous.4open.science/r/DFQ-84E6 and will be publicly available.
Feature Information Driven Position Gaussian Distribution Estimation for Tiny Object Detection
PDF ↗Tiny object detection remains challenging in spite of the success of generic detectors. The dramatic performance degradation of generic detectors on tiny objects is mainly due to the the weak representations of extremely limited pixels. To address this issue, we propose a plug-and-play architecture to enhance the extinguished regions. We for the first time exploit the regions to be enhanced from the perspective of pixel-wise amount of information. Specifically, we model the entire image pixels feature information by minimizing Information Entropy loss, generating an information map to attentively highlight weak activated regions in an unsupervised way. To effectively assist the above phase with more attention to tiny objects, we next introduce the Position Gaussian Distribution Map, explicitly modeled using a Gaussian Mixture distribution, where each Gaussian component's parameters depend on the position and size of object instance labels, serving as supervision for further feature enhancement. Taking the information map as prior knowledge guidance, we construct a multi-scale position gaussian distribution map prediction module, simultaneously modulating the information map and distribution map to focus on tiny objects during training. Extensive experiments on three public tiny object datasets demonstrate the superiority of our method over current state-of-the-art competitors.
Recent contrastive multimodal vision-language models like CLIP have demonstrated robust open-world semantic understanding, becoming the standard image backbones for vision-language applications. However, recent findings suggest high semantic similarity between well-trained unimodal encoders, which raises a key question: Are semantically similar embedding spaces separated only by simple projection transformations? To validate this, we propose a novel framework that aligns vision and language using frozen unimodal encoders. It involves selecting semantically similar encoders in the latent space, curating a concept-rich dataset of image-caption pairs, and training simple MLP projectors. We evaluated our approach on various tasks involving both strong unimodal vision (0-shot localization) and language encoders (multi-lingual, long context) and show that simple Projectors retain unimodal capabilities in joint embedding space. Furthermore, our best model, utilizing DINOv2 and All-Roberta-Large text encoder, achieves 76(%) accuracy on ImageNet with a 20-fold reduction in data and 65-fold reduction in compute requirements compared to multimodal alignment where models are trained from scratch. The proposed framework enhances the accessibility of multimodal model development while enabling flexible adaptation across diverse scenarios. Code and curated datasets are available at github.com/mayug/freeze-align.
Text-guided image manipulation has experienced notable advancement in recent years. In order to mitigate linguistic ambiguity, few-shot learning with visual examples has been applied for instructions that are underrepresented in the training set, or difficult to describe purely in language. However, learning from visual prompts requires strong reasoning capability, which diffusion models are struggling with. To address this issue, we introduce a novel multi-modal autoregressive model, dubbed InstaManip, that can instantly learn a new image manipulation operation from textual and visual guidance via in-context learning, and apply it to new query images. Specifically, we propose an innovative group self-attention mechanism to break down the in-context learning process into two separate stages -- learning and applying, which simplifies the complex problem into two easier tasks. We also introduce a relation regularization method to further disentangle image transformation features from irrelevant contents in exemplar images. Extensive experiments suggest that our method surpasses previous few-shot image manipulation models by a notable margin (>=19% in human evaluation). We also find our model can be further boosted by increasing the number or diversity of exemplar images.
Foundation models (FMs) such as CLIP have demonstrated impressive zero-shot performance across various tasks by leveraging large-scale, unsupervised pre-training. However, they often inherit harmful or unwanted knowledge from noisy internet-sourced datasets, compromising their reliability in real-world applications. Existing model unlearning methods either rely on access to pre-trained datasets or focus on coarse-grained unlearning (e.g., entire classes), leaving a critical gap for fine-grained unlearning. In this paper, we address the challenging scenario of selectively forgetting specific portions of knowledge within a class--without access to pre-trained data--while preserving the model's overall performance. We propose a novel three-stage approach that progressively unlearns targeted knowledge while mitigating over-forgetting. It consists of (1) a forgetting stage to fine-tune the CLIP on samples to be forgotten, (2) a reminding stage to restore performance on retained samples, and (3) a restoring stage to recover zero-shot capabilities using model souping. Additionally, we introduce knowledge distillation to handle the distribution disparity between forgetting/retaining samples and unseen pre-trained data. Extensive experiments on CIFAR-10, ImageNet-1K, and style datasets demonstrate that our approach effectively unlearns specific subgroups while maintaining strong zero-shot performance on semantically similar subgroups and other categories, significantly outperforming baseline unlearning methods, which lose effectiveness under the CLIP unlearning setting.
We propose a diffusion-based inverse rendering framework that decomposes a single RGB image into geometry, material, and lighting. Inverse rendering is inherently ill-posed, making it difficult to predict a single accurate solution. To address this challenge, recent generative model-based methods aim to present a range of possible solutions. However, finding a single accurate solution and generating diverse solutions can be conflicting. In this paper, we propose a channel-wise noise scheduling approach that allows a single diffusion model architecture to achieve two conflicting objectives. The resulting two diffusion models, trained with different channel-wise noise schedules, can predict a single highly accurate solution and present multiple possible solutions. The experimental results demonstrate the superiority of our two models in terms of both diversity and accuracy, which translates to enhanced performance in downstream applications such as object insertion and material editing.
Video Scene Graph Generation (VidSGG) aims to capture dynamic relationships among entities by sequentially analyzing video frames and integrating visual and semantic information. However, VidSGG is challenged by significant biases that skew predictions. To mitigate these biases, we propose a VIsual and Semantic Awareness (VISA) framework for unbiased VidSGG. VISA addresses visual bias through an innovative memory update mechanism that enhances object representations and concurrently reduces semantic bias by iteratively integrating object features with comprehensive semantic information derived from triplet relationships. This visual-semantics dual debiasing approach results in more unbiased representations of complex scene dynamics. Extensive experiments demonstrate the effectiveness of our method, where VISA outperforms existing unbiased VidSGG approaches by a substantial margin (e.g., +13.1% improvement in mR@20 and mR@50 for the SGCLS task under Semi Constraint).