We investigate compositional structures in data embeddings from pre-trained vision-language models (VLMs). Traditionally, compositionality has been associated with algebraic operations on embeddings of words from a pre-existing vocabulary. In contrast, we seek to approximate representations from an encoder as combinations of a smaller set of vectors in the embedding space. These vectors can be seen as "ideal words" for generating concepts directly within embedding space of the model. We first present a framework for understanding compositional structures from a geometric perspective. We then explain what these compositional structures entail probabilistically in the case of VLM embeddings, providing intuitions for why they arise in practice. Finally, we empirically explore these structures in CLIP's embeddings and we evaluate their usefulness for solving different vision-language tasks such as classification, debiasing, and retrieval. Our results show that simple linear algebraic operations on embedding vectors can be used as compositional and interpretable methods for regulating the behavior of VLMs.
论文检索
输入标题、作者或关键词，从 1,144 篇学术成果中精准定位
Large vision-language models, such as CLIP, learn robust representations of text and images, facilitating advances in many downstream tasks, including zero-shot classification and text-to-image generation. However, these models have several well-documented limitations. They fail to encapsulate compositional concepts, such as counting. To the best of our knowledge, this work is the first to extend CLIP to handle object counting. We introduce a simple yet effective method to improve the quantitative understanding of vision-language models, while maintaining their overall performance on common benchmarks. Our method automatically augments image captions to create hard negative samples that differ from the original captions by only the number of objects. For example, an image of three dogs can be contrasted with the negative caption "Six dogs playing in the yard". A dedicated loss encourages discrimination between the correct caption and its negative variant. In addition, we introduce CountBench, a new benchmark for evaluating a model's understanding of object counting, and demonstrate significant improvement over baseline models on this task. Furthermore, we leverage our improved CLIP representations for text-conditioned image generation, and show that our model can produce specific counts of objects more reliably than existing ones.
We address the problem of network calibration adjusting miscalibrated confidences of deep neural networks. Many approaches to network calibration adopt a regularization-based method that exploits a regularization term to smooth the miscalibrated confidences. Although these approaches have shown the effectiveness on calibrating the networks, there is still a lack of understanding on the underlying principles of regularization in terms of network calibration. We present in this paper an in-depth analysis of existing regularization-based methods, providing a better understanding on how they affect to network calibration. Specifically, we have observed that 1) the regularization-based methods can be interpreted as variants of label smoothing, and 2) they do not always behave desirably. Based on the analysis, we introduce a novel loss function, dubbed ACLS, that unifies the merits of existing regularization methods, while avoiding the limitations. We show extensive experimental results for image classification and semantic segmentation on standard benchmarks, including CIFAR10, Tiny-ImageNet, ImageNet, and PASCAL VOC, demonstrating the effectiveness of our loss function.
Masked Spatio-Temporal Structure Prediction for Self-supervised Learning on Point Cloud Videos
PDF ↗Recently, the community has made tremendous progress in developing effective methods for point cloud video understanding that learn from massive amounts of labeled data. However, annotating point cloud videos is usually notoriously expensive. Moreover, training via one or only a few traditional tasks (e.g., classification) may be insufficient to learn subtle details of the spatio-temporal structure existing in point cloud videos. In this paper, we propose a Masked Spatio-Temporal Structure Prediction (MaST-Pre) method to capture the structure of point cloud videos without human annotations. MaST-Pre is based on spatio-temporal point-tube masking and consists of two self-supervised learning tasks. First, by reconstructing masked point tubes, our method is able to capture the appearance information of point cloud videos. Second, to learn motion, we propose a temporal cardinality difference prediction task that estimates the change in the number of points within a point tube. In this way, MaST-Pre is forced to model the spatial and temporal structure in point cloud videos. Extensive experiments on MSRAction-3D, NTU-RGBD, NvGesture, and SHREC'17 demonstrate the effectiveness of the proposed method.
FishNet: A Large-scale Dataset and Benchmark for Fish Recognition, Detection, and Functional Trait Prediction
PDF ↗Aquatic species are essential components of the world's ecosystem, and the preservation of aquatic biodiversity is crucial for maintaining proper ecosystem functioning. Unfortunately, increasing anthropogenic pressures such as overfishing, climate change, and coastal development pose significant threats to aquatic biodiversity. To address this challenge, it is necessary to design an automatic aquatic species monitoring systems that can help researchers and policymakers better understand changes in aquatic ecosystems and take appropriate actions to preserve biodiversity. However, the development of such systems is impeded by a lack of large-scale diverse aquatic species datasets. Existing aquatic species recognition datasets generally have a limited number of species, nor do they provide functional trait data, and so have only narrow potential for application. To address the need for generalized systems that can recognize, locate, and predict a wide array of species and their functional traits, we present FishNet, a large-scale diverse dataset containing 94,532 meticulously organized images from 17,357 aquatic species, organized according to aquatic biological taxonomy (order, family, genus, and species). We further build three benchmarks, i.e., fish classification, fish detection, and functional trait prediction, inspired by ecological research needs, to facilitate the development of aquatic species recognition systems, and promote further research in the field of aquatic ecology. Our FishNet dataset has the potential to encourage the development of more accurate and effective tools for the monitoring and protection of aquatic ecosystems, and hence take effective action toward the conservation of our planet's aquatic biodiversity. Our dataset and code will be released at https://fishnet-2023.github.io/.
One-Shot Recognition of Any Material Anywhere Using Contrastive Learning with Physics-Based Rendering
PDF ↗Visual recognition of materials and their states is essential for understanding the world, from determining whether food is cooked, metal is rusted, or a chemical reaction has occurred. However, current image recognition methods are limited to specific classes and properties and can't handle the vast number of material states in the world. To address this, we present MatSim: the first dataset and benchmark for computer vision-based recognition of similarities and transitions between materials and textures, focusing on identifying any material under any conditions using one or a few examples. The dataset contains synthetic and natural images. Synthetic images were rendered using giant collections of textures, objects, and environments generated by computer graphics artists. We use mixtures and gradual transitions between materials to allow the system to learn cases with smooth transitions between states (like gradually cooked food). We also render images with materials inside transparent containers to support beverage and chemistry lab use cases. We use this dataset to train a Siamese net that identifies the same material in different objects, mixtures, and environments. The descriptor generated by this net can be used to identify the states of materials and their subclasses using a single image. We also present the first few-shot material recognition benchmark with natural images from a wide range of fields, including the state of foods and beverages, types of grounds, and many other use cases. We show that a net trained on the MatSim synthetic dataset outperforms state-of-the-art models like Clip on the benchmark and also achieves good results on other unsupervised material classification tasks. Dataset, generation code and trained models have been made available at: https://github.com/ZuseZ4/MatSim-Dataset-Generator-Scripts-And-Neural-net
There has been a longstanding belief that generation can facilitate a true understanding of visual data. In line with this, we revisit generatively pre-training visual representations in light of recent interest in denoising diffusion models. While directly pre-training with diffusion models does not produce strong representations, we condition diffusion models on masked input and formulate diffusion models as masked autoencoders (DiffMAE). Our approach is capable of (i) serving as a strong initialization for downstream recognition tasks, (ii) conducting high-quality image inpainting, and (iii) being effortlessly extended to video where it produces state-of-the-art classification accuracy. We further perform a comprehensive study on the pros and cons of design choices and build connections between diffusion models and masked autoencoders.
Video relation grounding (VRG) is a significant and challenging problem in the domains of cross-modal learning and video understanding. In this study, we introduce a novel approach called inverse compositional learning (ICL) for weakly-supervised video relation grounding. Our approach represents relations at both the holistic and partial levels, formulating VRG as a joint optimization problem that encompasses reasoning at both levels. For holistic-level reasoning, we propose an inverse attention mechanism and a compositional encoder to generate compositional relevance features. Additionally, we introduce an inverse loss to evaluate and learn the relevance between visual features and relation features. At the partial-level reasoning, we introduce a grounding by classification scheme. By leveraging the learned holistic-level features and partial-level features, we train the entire model in an end-to-end manner. We conduct evaluations on two challenging datasets and demonstrate the substantial superiority of our proposed method over state-of-the-art methods. Extensive ablation studies confirm the effectiveness of our approach.
We propose Spatio-temporal Crop Aggregation for video representation LEarning (SCALE), a novel method that enjoys high scalability at both training and inference time. Our model builds long-range video features by learning from sets of video clip-level features extracted with a pre-trained backbone. To train the model, we propose a self-supervised objective consisting of masked clip feature predictions. We apply sparsity to both the input, by extracting a random set of video clips, and to the loss function, by only reconstructing the sparse inputs. Moreover, we use dimensionality reduction by working in the latent space of a pre-trained backbone applied to single video clips. These techniques make our method not only extremely efficient to train but also highly effective in transfer learning. We demonstrate that our video representation yields state-of-the-art performance with linear, nonlinear, and k-NN probing on common action classification and video understanding datasets.
A number of computer vision deep regression approaches report improved results when adding a classification loss to the regression loss. Here, we explore why this is useful in practice and when it is beneficial. To do so, we start from precisely controlled dataset variations and data samplings and find that the effect of adding a classification loss is the most pronounced for regression with imbalanced data. We explain these empirical findings by formalizing the relation between the balanced and imbalanced regression losses. Finally, we show that our findings hold on two real imbalanced image datasets for depth estimation (NYUD2-DIR), and age estimation (IMDB-WIKI-DIR), and on the problem of imbalanced video progress prediction (Breakfast). Our main takeaway is: for a regression task, if the data sampling is imbalanced, then add a classification loss.
Understanding verbs is crucial to modelling how people and objects interact with each other and the environment through space and time. Recently, state-of-the-art video-language models based on CLIP have been shown to have limited verb understanding and to rely extensively on nouns, restricting their performance in real-world video applications that require action and temporal understanding. In this work, we improve verb understanding for CLIP-based video-language models by proposing a new Verb-Focused Contrastive (VFC) framework. This consists of two main components: (1) leveraging pretrained large language models (LLMs) to create hard negatives for cross-modal contrastive learning, together with a calibration strategy to balance the occurrence of concepts in positive and negative pairs; and (2) enforcing a fine-grained, verb phrase alignment loss. Our method achieves state-of-the-art results for zero-shot performance on three downstream tasks that focus on verb understanding, including video-text matching, video question-answering and video classification; while maintaining performance on noun-focused settings. To the best of our knowledge, this is the first work which proposes a method to alleviate the verb understanding problem, and does not simply highlight it. Our code is publicly available.
A neural network trained on a classification dataset often exhibits a higher vector norm of hidden layer features for in-distribution (ID) samples, while producing relatively lower norm values on unseen instances from out-of-distribution (OOD). Despite this intriguing phenomenon being utilized in many applications, the underlying cause has not been thoroughly investigated. In this study, we demystify this very phenomenon by scrutinizing the discriminative structures concealed in the intermediate layers of a neural network. Our analysis leads to the following discoveries: (1) The feature norm is a confidence value of a classifier hidden in the network layer, specifically its maximum logit. Hence, the feature norm distinguishes OOD from ID in the same manner that a classifier confidence does. (2) The feature norm is class-agnostic, thus it can detect OOD samples across diverse discriminative models. (3) The conventional feature norm fails to capture the deactivation tendency of hidden layer neurons, which may lead to misidentification of ID samples as OOD instances. To resolve this drawback, we propose a novel negative-aware norm (NAN) that can capture both the activation and deactivation tendencies of hidden layer neurons. We conduct extensive experiments on NAN, demonstrating its efficacy and compatibility with existing OOD detectors, as well as its capability in label-free environments.
Computer vision models have known performance disparities across attributes such as gender and skin tone. This means during tasks such as classification and detection, model performance differs for certain classes based on the demographics of the people in the image. These disparities have been shown to exist, but until now there has not been a unified approach to measure these differences for common use-cases of computer vision models. We present a new benchmark named FACET (FAirness in Computer Vision EvaluaTion), a large, publicly available evaluation set of 32k images for some of the most common vision tasks - image classification, object detection and segmentation. For every image in FACET, we hired expert reviewers to manually annotate person-related attributes such as perceived skin tone and hair type, manually draw bounding boxes and label fine-grained person-related classes such as disk jockey or guitarist. In addition, we use FACET to benchmark state-of-the-art vision models and present a deeper understanding of potential performance disparities and challenges across sensitive demographic attributes. With the exhaustive annotations collected, we probe models using single demographics attributes as well as multiple attributes using an intersectional approach (e.g. hair color and perceived skin tone). Our results show that classification, detection, segmentation, and visual grounding models exhibit performance disparities across demographic attributes and intersections of attributes. These harms suggest that not all people represented in datasets receive fair and equitable treatment in these vision tasks. We hope current and future results using our benchmark will contribute to fairer, more robust vision models. FACET is available publicly at https://facet.metademolab.com
Personalized Federated Learning (PFL) represents a promising solution for decentralized learning in heterogeneous data environments. Partial model personalization has been proposed to improve the efficiency of PFL by selectively updating local model parameters instead of aggregating all of them. However, previous work on partial model personalization has mainly focused on Convolutional Neural Networks (CNNs), leaving a gap in understanding how it can be applied to other popular models such as Vision Transformers (ViTs). In this work, we investigate where and how to partially personalize a ViT model. Specifically, we empirically evaluate the sensitivity to data distribution of each type of layer. Based on the insights that the self-attention layer and the classification head are the most sensitive parts of a ViT, we propose a novel approach called FedPerfix, which leverages plugins to transfer information from the aggregated model to the local client as a personalization. Finally, we evaluate the proposed approach on CIFAR-100, OrganAMNIST, and Office-Home datasets and demonstrate its effectiveness in improving the model's performance compared to several advanced PFL methods.
Pre-trained point cloud models have found extensive applications in 3D understanding tasks like object classification and part segmentation. However, the prevailing strategy of full fine-tuning in downstream tasks leads to large per-task storage overhead for model parameters, which limits the efficiency when applying large-scale pre-trained models. Inspired by the recent success of visual prompt tuning (VPT), this paper attempts to explore prompt tuning on pre-trained point cloud models, to pursue an elegant balance between performance and parameter efficiency. We find while instance-agnostic static prompting, e.g. VPT, shows some efficacy in downstream transfer, it is vulnerable to the distribution diversity caused by various types of noises in real-world point cloud data. To conquer this limitation, we propose a novel Instance-aware Dynamic Prompt Tuning (IDPT) strategy for pre-trained point cloud models. The essence of IDPT is to develop a dynamic prompt generation module to perceive semantic prior features of each point cloud instance and generate adaptive prompt tokens to enhance the model's robustness. Notably, extensive experiments demonstrate that IDPT outperforms full fine-tuning in most tasks with a mere 7% of the trainable parameters, providing a promising solution to parameter-efficient learning for pre-trained point cloud models. Code is available at https://github.com/zyh16143998882/ICCV23-IDPT.
Frequency analysis is useful for understanding the mechanisms of representation learning in neural networks (NNs). Most research in this area focuses on the learning dynamics of NNs for regression tasks, while little for classification. This study empirically investigates the latter and expands the understanding of frequency shortcuts. First, we perform experiments on synthetic datasets, designed to have a bias in different frequency bands. Our results demonstrate that NNs tend to find simple solutions for classification, and what they learn first during training depends on the most distinctive frequency characteristics, which can be either low- or high-frequencies. Second, we confirm this phenomenon on natural images. We propose a metric to measure class-wise frequency characteristics and a method to identify frequency shortcuts. The results show that frequency shortcuts can be texture-based or shape-based, depending on what best simplifies the objective. Third, we validate the transferability of frequency shortcuts on out-of-distribution (OOD) test sets. Our results suggest that frequency shortcuts can be transferred across datasets and cannot be fully avoided by larger model capacity and data augmentation. We recommend that future research should focus on effective training schemes mitigating frequency shortcut learning. Codes and data are available at https://github.com/nis-research/nn-frequency-shortcuts.
The representation of multimedia content experienced tremendous advances in the last decades. Mainly supported by deep learning models, impressive results have been obtained. However, despite such advances in representation, the definition of similarity has been neglected. Effectively computing the similarity between representations remains a challenge. Traditional distance functions, such as the Euclidean distance, are not able to properly consider the relevant similarity information encoded in the dataset manifold. In fact, manifolds are essential to perception in many scenarios, such that exploiting the underlying structure of dataset manifolds plays a central role in multimedia content understanding and retrieval. In this paper, we present a framework for unsupervised distance learning which provides easy and uniform access to methods capable of considering the dataset manifold for redefining similarity. Such methods perform context-sensitive similarity learning based on more global measures, capable of improving the effectiveness of retrieval and machine learning tasks. The framework can use distance, similarity, or ranking information both as input and output and compute traditional retrieval effectiveness measures. Implemented as a wrapper in Python, the framework allows integration with a large number of Python libraries while keeping a back-end in C++ for efficiency. The paper also discusses diverse applications of the methods available in the pyUDLF framework, including image re-ranking, video retrieval, person re-ID, and pre-processing of distance measurements for clustering and classification.
AI has made significant progress in solving math problems, but geometry problems remain challenging due to their reliance on both text and diagrams. In the text description, symbolic characters such as "ABC" often serve as a bridge to connect the corresponding diagram. However, by simply tokenizing symbolic characters into individual letters (e.g., 'A', 'B' and 'C'), existing works fail to study them explicitly and thus lose the semantic relationship with the diagram. In this paper, we develop a symbolic character-aware model to fully explore the role of these characters in both text and diagram understanding and optimize the model under a multi-modal reasoning framework. In the text encoder, we propose merging individual symbolic characters to form one semantic unit along with geometric information from the corresponding diagram. For the diagram encoder, we pre-train it under a multi-label classification framework with the symbolic characters as labels. In addition, we enhance the geometry diagram understanding ability via a self-supervised learning method under the masked image modeling auxiliary task. By integrating the proposed model into a general encoder-decoder pipeline for solving geometry problems, we demonstrate its superiority on two benchmark datasets, including GeoQA and Geometry3K, with extensive experiments. Specifically, on GeoQA, the question-solving accuracy is increased from 60.0% to 64.1%, achieving a new state-of-the-art accuracy; on Geometry3K, we reduce the question average solving steps from 6.9 down to 6.0 with marginally higher solving accuracy.
Exploring Shape Embedding for Cloth-Changing Person Re-Identification via 2D-3D Correspondences
PDF ↗Cloth-Changing Person Re-Identification (CC-ReID) is a common and realistic problem since fashion constantly changes over time and people's aesthetic preferences are not set in stone. While most existing cloth-changing ReID methods focus on learning cloth-agnostic identity representations from coarse semantic cues (e.g. silhouettes and part segmentation maps), they neglect the continuous shape distributions at the pixel level. In this paper, we propose Continuous Surface Correspondence Learning (CSCL), a new shape embedding paradigm for cloth-changing ReID. CSCL establishes continuous correspondences between a 2D image plane and a canonical 3D body surface via pixel-to-vertex classification, which naturally aligns a person image to the surface of a 3D human model and simultaneously obtains pixel-wise surface embeddings. We further extract fine-grained shape features from the learned surface embeddings and then integrate them with global RGB features via a carefully designed cross-modality fusion module. The shape embedding paradigm based on 2D-3D correspondences remarkably enhances the model's global understanding of human body shape. To promote the study of ReID under clothing change, we construct 3D Dense Persons (DP3D), which is the first large-scale cloth-changing ReID dataset that provides densely annotated 2D-3D correspondences and a precise 3D mesh for each person image, while containing diverse cloth-changing cases over all four seasons. Experiments on both cloth-changing and cloth-consistent ReID benchmarks validate the effectiveness of our method.
Most existing methods for image emotion classification extract features directly from images supervised by a single emotional label. However, this approach has a limitation known as the affective gap which restricts the capability of these features as they do not always align with the emotions perceived by users. To effectively bridge the affective gap, this paper proposes a visual content understanding network inspired by the human staged emotion perception process. The proposed network is comprised of three perception modules designed to extract multi-level information. Firstly, an entity perception module extracts entities from images. Secondly, an attribute perception module extracts the attribute content of each entity. Thirdly, an emotion perception module extracts emotion features based on both the entity and attribute information. We generate pseudo-labels of entities and attributes through image segmentation and vision-language models to provide auxiliary guidance for network learning. The progressive entity and attribute understanding enable the network to hierarchically extract semantic-level features for emotion analysis. Extensive experiments demonstrate that our progressive learning network achieves superior performance on various benchmark datasets for image emotion classification.