NeurIPS 2024 Papers

Recently, there have been explorations of generalist segmentation models that can effectively tackle a variety of image segmentation tasks within a unified in-context learning framework. However, these methods still struggle with task ambiguity in in-context segmentation, as not all in-context examples can accurately convey the task information. In order to address this issue, we present SINE, a simple image $\textbf{S}$egmentation framework utilizing $\textbf{in}$-context $\textbf{e}$xamples. Our approach leverages a Transformer encoder-decoder structure, where the encoder provides high-quality image representations, and the decoder is designed to yield multiple task-specific output masks to eliminate task ambiguity effectively. Specifically, we introduce an In-context Interaction module to complement in-context information and produce correlations between the target image and the in-context example and a Matching Transformer that uses fixed matching and a Hungarian algorithm to eliminate differences between different tasks. In addition, we have further perfected the current evaluation system for in-context image segmentation, aiming to facilitate a holistic appraisal of these models. Experiments on various segmentation tasks show the effectiveness of the proposed method.

Recent vision foundation models can extract universal representations and show impressive abilities in various tasks. However, their application on object detection is largely overlooked, especially without fine-tuning them. In this work, we show that frozen foundation models can be a versatile feature enhancer, even though they are not pre-trained for object detection. Specifically, we explore directly transferring the high-level image understanding of foundation models to detectors in the following two ways. First, the class token in foundation models provides an in-depth understanding of the complex scene, which facilitates decoding object queries in the detector's decoder by providing a compact context. Additionally, the patch tokens in foundation models can enrich the features in the detector's encoder by providing semantic details. Utilizing frozen foundation models as plug-and-play modules rather than the commonly used backbone can significantly enhance the detector's performance while preventing the problems caused by the architecture discrepancy between the detector's backbone and the foundation model. With such a novel paradigm, we boost the SOTA query-based detector DINO from 49.0% AP to 51.9% AP (+2.9% AP) and further to 53.8% AP (+4.8% AP) by integrating one or two foundation models respectively, on the COCO validation set after training for 12 epochs with R50 as the detector's backbone. Code will be available.

The Area Under the ROC Curve (AUC) is a well-known metric for evaluating instance-level long-tail learning problems. In the past two decades, many AUC optimization methods have been proposed to improve model performance under long-tail distributions. In this paper, we explore AUC optimization methods in the context of pixel-level long-tail semantic segmentation, a much more complicated scenario. This task introduces two major challenges for AUC optimization techniques. On one hand, AUC optimization in a pixel-level task involves complex coupling across loss terms, with structured inner-image and pairwise inter-image dependencies, complicating theoretical analysis. On the other hand, we find that mini-batch estimation of AUC loss in this case requires a larger batch size, resulting in an unaffordable space complexity. To address these issues, we develop a pixel-level AUC loss function and conduct a dependency-graph-based theoretical analysis of the algorithm's generalization ability. Additionally, we design a Tail-Classes Memory Bank (T-Memory Bank) to manage the significant memory demand. Finally, comprehensive experiments across various benchmarks confirm the effectiveness of our proposed AUCSeg method. The code is available at https://github.com/boyuh/AUCSeg.

Understanding egocentric human-object interaction (HOI) is a fundamental aspect of human-centric perception, facilitating applications like AR/VR and embodied AI. For the egocentric HOI, in addition to perceiving semantics e.g., ''what'' interaction is occurring, capturing ''where'' the interaction specifically manifests in 3D space is also crucial, which links the perception and operation. Existing methods primarily leverage observations of HOI to capture interaction regions from an exocentric view. However, incomplete observations of interacting parties in the egocentric view introduce ambiguity between visual observations and interaction contents, impairing their efficacy. From the egocentric view, humans integrate the visual cortex, cerebellum, and brain to internalize their intentions and interaction concepts of objects, allowing for the pre-formulation of interactions and making behaviors even when interaction regions are out of sight. In light of this, we propose harmonizing the visual appearance, head motion, and 3D object to excavate the object interaction concept and subject intention, jointly inferring 3D human contact and object affordance from egocentric videos. To achieve this, we present EgoChoir, which links object structures with interaction contexts inherent in appearance and head motion to reveal object affordance, further utilizing it to model human contact. Additionally, a gradient modulation is employed to adopt appropriate clues for capturing interaction regions across various egocentric scenarios. Moreover, 3D contact and affordance are annotated for egocentric videos collected from Ego-Exo4D and GIMO to support the task. Extensive experiments on them demonstrate the effectiveness and superiority of EgoChoir.

State Space Models (SSMs) have the advantage of keeping linear computational complexity compared to attention modules in transformers, and have been applied to vision tasks as a new type of powerful vision foundation model. Inspired by the observations that the final prediction in vision transformers (ViTs) is only based on a subset of most informative tokens, we take the novel step of enhancing the efficiency of SSM-based vision models through token-based pruning. However, direct applications of existing token pruning techniques designed for ViTs fail to deliver good performance, even with extensive fine-tuning. To address this issue, we revisit the unique computational characteristics of SSMs and discover that naive application disrupts the sequential token positions. This insight motivates us to design a novel and general token pruning method specifically for SSM-based vision models. We first introduce a pruning-aware hidden state alignment method to stabilize the neighborhood of remaining tokens for performance enhancement. Besides, based on our detailed analysis, we propose a token importance evaluation method adapted for SSM models, to guide the token pruning. With efficient implementation and practical acceleration methods, our method brings actual speedup. Extensive experiments demonstrate that our approach can achieve significant computation reduction with minimal impact on performance across different tasks. Notably, we achieve 81.7\% accuracy on ImageNet with a 41.6\% reduction in the FLOPs for pruned PlainMamba-L3. Furthermore, our work provides deeper insights into understanding the behavior of SSM-based vision models for future research.

The goal of this paper is to improve the generality and accuracy of open-vocabulary object counting in images. To improve the generality, we repurpose an open-vocabulary detection foundation model (GroundingDINO) for the counting task, and also extend its capabilities by introducing modules to enable specifying the target object to count by visual exemplars. In turn, these new capabilities -- being able to specify the target object by multi-modalites (text and exemplars) -- lead to an improvement in counting accuracy. We make three contributions: First, we introduce the first open-world counting model, CountGD, where the prompt can be specified by a text description or visual exemplars or both; Second, we show that the performance of the model significantly improves the state of the art on multiple counting benchmarks -- when using text only, CountGD outperforms all previous text-only works, and when using both text and visual exemplars, we outperform all previous models; Third, we carry out a preliminary study into different interactions between the text and visual exemplar prompts, including the cases where they reinforce each other and where one restricts the other. The code and an app to test the model are available at https://www.robots.ox.ac.uk/vgg/research/countgd/.

Vision-Language Models seamlessly discriminate among arbitrary semantic categories, yet they still suffer from poor generalization when presented with challenging examples. For this reason, Episodic Test-Time Adaptation (TTA) strategies have recently emerged as powerful techniques to adapt VLMs in the presence of a single unlabeled image. The recent literature on TTA is dominated by the paradigm of prompt tuning by Marginal Entropy Minimization, which, relying on online backpropagation, inevitably slows down inference while increasing memory. In this work, we theoretically investigate the properties of this approach and unveil that a surprisingly strong TTA method lies dormant and hidden within it. We term this approach ZERO (TTA with “zero” temperature), whose design is both incredibly effective and frustratingly simple: augment N times, predict, retain the most confident predictions, and marginalize after setting the Softmax temperature to zero. Remarkably, ZERO requires a single batched forward pass through the vision encoder only and no backward passes. We thoroughly evaluate our approach following the experimental protocol established in the literature and show that ZERO largely surpasses or compares favorably w.r.t. the state-of-the-art while being almost 10× faster and 13× more memory friendly than standard Test-Time Prompt Tuning. Thanks to its simplicity and comparatively negligible computation, ZERO can serve as a strong baseline for future work in this field. Code will be available.

As a widely explored multi-modal task, Temporal Sentence Grounding in videos (TSG) endeavors to retrieve a specific video segment matched with a given query text from a video. The traditional paradigm for TSG generally assumes that relevant segments always exist within a given video. However, this assumption is restrictive and unrealistic in real-world applications where the existence of a query-related segment is uncertain, easily resulting in erroneous grounding. Motivated by the research gap and practical application, this paper introduces a new task, named Temporal Sentence Grounding with Relevance Feedback (TSG-RF) in videos, which accommodates the possibility that a video may or may not include a segment related to the query. This task entails localizing precise video segments that semantically align with the query text when such content is present, while delivering definitive feedback on the non-existence of related segments when absent. Moreover, we propose a novel Relation-aware Temporal Sentence Grounding (RaTSG) network for addressing this challenging task. This network first reformulates the TSG-RF task as a foreground-background detection problem by investigating whether the query-related semantics exist in both frame and video levels. Then, a multi-granularity relevance discriminator is exploited to produce precise video-query relevance feedback and a relation-aware segment grounding module is employed to selectively conduct the grounding process, dynamically adapting to the presence or absence of query-related segments in videos. To validate our RaTSG network, we reconstruct two popular TSG datasets, establishing a rigorous benchmark for TSG-RF. Experimental results demonstrate the effectiveness of our proposed RaTSG for the TSG-RF task. Our source code is available at https://github.com/HuiGuanLab/RaTSG.

Surface parameterization plays an essential role in numerous computer graphics and geometry processing applications. Traditional parameterization approaches are designed for high-quality meshes laboriously created by specialized 3D modelers, thus unable to meet the processing demand for the current explosion of ordinary 3D data. Moreover, their working mechanisms are typically restricted to certain simple topologies, thus relying on cumbersome manual efforts (e.g., surface cutting, part segmentation) for pre-processing. In this paper, we introduce the Flatten Anything Model (FAM), an unsupervised neural architecture to achieve global free-boundary surface parameterization via learning point-wise mappings between 3D points on the target geometric surface and adaptively-deformed UV coordinates within the 2D parameter domain. To mimic the actual physical procedures, we ingeniously construct geometrically-interpretable sub-networks with specific functionalities of surface cutting, UV deforming, unwrapping, and wrapping, which are assembled into a bi-directional cycle mapping framework. Compared with previous methods, our FAM directly operates on discrete surface points without utilizing connectivity information, thus significantly reducing the strict requirements for mesh quality and even applicable to unstructured point cloud data. More importantly, our FAM is fully-automated without the need for pre-cutting and can deal with highly-complex topologies, since its learning process adaptively finds reasonable cutting seams and UV boundaries. Extensive experiments demonstrate the universality, superiority, and inspiring potential of our proposed neural surface parameterization paradigm. Our code is available at https://github.com/keeganhk/FlattenAnything.

Despite the substantial progress of novel view synthesis, existing methods, either based on the Neural Radiance Fields (NeRF) or more recently 3D Gaussian Splatting (3DGS), suffer significant degradation when the input becomes sparse. Numerous efforts have been introduced to alleviate this problem, but they still struggle to synthesize satisfactory results efficiently, especially in the large scene. In this paper, we propose SCGaussian, a Structure Consistent Gaussian Splatting method using matching priors to learn 3D consistent scene structure. Considering the high interdependence of Gaussian attributes, we optimize the scene structure in two folds: rendering geometry and, more importantly, the position of Gaussian primitives, which is hard to be directly constrained in the vanilla 3DGS due to the non-structure property. To achieve this, we present a hybrid Gaussian representation. Besides the ordinary non-structure Gaussian primitives, our model also consists of ray-based Gaussian primitives that are bound to matching rays and whose optimization of their positions is restricted along the ray. Thus, we can utilize the matching correspondence to directly enforce the position of these Gaussian primitives to converge to the surface points where rays intersect. Extensive experiments on forward-facing, surrounding, and complex large scenes show the effectiveness of our approach with state-of-the-art performance and high efficiency. Code is available at https://github.com/prstrive/SCGaussian.

Tracking points in video frames is essential for understanding video content. However, the task is fundamentally hindered by the computation demands for brute-force correspondence matching across the frames. As the current models down-sample the frame resolutions to mitigate this challenge, they fall short in accurately representing point trajectories due to information truncation. Instead, we address the challenge by pruning the search space for point tracking and let the model process only the important regions of the frames without down-sampling. Our first key idea is to identify the object instance and its trajectory over the frames, then prune the regions of the frame that do not contain the instance. Concretely, to estimate the instance’s trajectory, we track a group of points on the instance and aggregate their motion trajectories. Furthermore, to deal with the occlusions in complex scenes, we propose to compensate for the occluded points while tracking. To this end, we introduce a unified framework that jointly performs point tracking and segmentation, providing synergistic effects between the two tasks. For example, the segmentation results enable a tracking model to avoid the occluded points referring to the instance mask, and conversely, the improved tracking results can help to produce more accurate segmentation masks. Our framework can be easily incorporated with various tracking models, and we demonstrate its efficacy for enhanced point tracking throughout extensive experiments. For example, on the recent TAP-Vid benchmark, our framework consistently improves all baselines, e.g., up to 13.5% improvement on the average Jaccard metric.

Language-image pre-training is an effective technique for learning powerful representations in general domains. However, when directly turning to person representation learning, these general pre-training methods suffer from unsatisfactory performance. The reason is that they neglect critical person-related characteristics, i.e., fine-grained attributes and identities. To address this issue, we propose a novel language-image pre-training framework for person representation learning, termed PLIP. Specifically, we elaborately design three pretext tasks: 1) Text-guided Image Colorization, aims to establish the correspondence between the person-related image regions and the fine-grained color-part textual phrases. 2) Image-guided Attributes Prediction, aims to mine fine-grained attribute information of the person body in the image; and 3) Identity-based Vision-Language Contrast, aims to correlate the cross-modal representations at the identity level rather than the instance level. Moreover, to implement our pre-train framework, we construct a large-scale person dataset with image-text pairs named SYNTH-PEDES by automatically generating textual annotations. We pre-train PLIP on SYNTH-PEDES and evaluate our models by spanning downstream person-centric tasks. PLIP not only significantly improves existing methods on all these tasks, but also shows great ability in the zero-shot and domain generalization settings. The code, dataset and weight will be made publicly available.

Graph autoencoders (Graph-AEs) learn representations of given graphs by aiming to accurately reconstruct them. A notable application of Graph-AEs is graph-level anomaly detection (GLAD), whose objective is to identify graphs with anomalous topological structures and/or node features compared to the majority of the graph population. Graph-AEs for GLAD regard a graph with a high mean reconstruction error (i.e. mean of errors from all node pairs and/or nodes) as anomalies. Namely, the methods rest on the assumption that they would better reconstruct graphs with similar characteristics to the majority. We, however, report non-trivial counter-examples, a phenomenon we call reconstruction flip, and highlight the limitations of the existing Graph-AE-based GLAD methods. Specifically, we empirically and theoretically investigate when this assumption holds and when it fails. Through our analyses, we further argue that, while the reconstruction errors for a given graph are effective features for GLAD, leveraging the multifaceted summaries of the reconstruction errors, beyond just mean, can further strengthen the features. Thus, we propose a novel and simple GLAD method, named MUSE. The key innovation of MUSE involves taking multifaceted summaries of reconstruction errors as graph features for GLAD. This surprisingly simple method obtains SOTA performance in GLAD, performing best overall among 14 methods across 10 datasets.

Existing parameter-efficient fine-tuning (PEFT) methods have achieved significant success on vision transformers (ViTs) adaptation by improving parameter efficiency. However, the exploration of enhancing inference efficiency during adaptation remains underexplored. This limits the broader application of pre-trained ViT models, especially when the model is computationally extensive. In this paper, we propose Dynamic Tuning (DyT), a novel approach to improve both parameter and inference efficiency for ViT adaptation. Specifically, besides using the lightweight adapter modules, we propose a token dispatcher to distinguish informative tokens from less important ones, allowing the latter to dynamically skip the original block, thereby reducing the redundant computation during inference. Additionally, we explore multiple design variants to find the best practice of DyT. Finally, inspired by the mixture-of-experts (MoE) mechanism, we introduce an enhanced adapter to further boost the adaptation performance. We validate DyT across various tasks, including image/video recognition and semantic segmentation. For instance, DyT achieves superior performance compared to existing PEFT methods while evoking only 71% of their FLOPs on the VTAB-1K benchmark.

The growing interest in vision-language models (VLMs) has been driven by improvements in large language models and vision transformers. Despite the abundance of literature on this subject, we observe that critical decisions regarding the design of VLMs are often not justified. We argue that these unsupported decisions impede progress in the field by making it difficult to identify which choices improve model performance. To address this issue, we conduct extensive experiments around pre-trained models, architecture choice, data, and training methods. Our consolidation of findings includes the development of Idefics2, an efficient foundational VLM of 8 billion parameters. Idefics2 achieves state-of-the-art performance within its size category across various multimodal benchmarks, and is often on par with models four times its size. We release the model (base, instructed, and chat) along with the datasets created for its training.

Multimodal Large Language Models (MLLMs) have gained significant attention due to their impressive capabilities in multimodal understanding. However, existing methods rely heavily on extensive modal-specific pretraining and joint-modal tuning, leading to significant computational burdens when expanding to new modalities. In this paper, we propose \textbf{PathWeave}, a flexible and scalable framework with modal-\textbf{path} s\textbf{w}itching and \textbf{e}xp\textbf{a}nsion abilities that enables MLLMs to continually \textbf{ev}olve on modalities for $\mathbb{X}$-modal reasoning. We leverage the concept of Continual Learning and develop an incremental training strategy atop pre-trained MLLMs, enabling their expansion to new modalities using uni-modal data, without executing joint-modal pretraining. In detail, a novel Adapter-in-Adapter (AnA) framework is introduced, in which uni-modal and cross-modal adapters are seamlessly integrated to facilitate efficient modality alignment and collaboration. Additionally, an MoE-based gating module is applied between two types of adapters to further enhance the multimodal interaction. To investigate the proposed method, we establish a challenging benchmark called \textbf{C}ontinual \textbf{L}earning of \textbf{M}odality (MCL), which consists of high-quality QA data from five distinct modalities: image, video, \textcolor{black}{audio, depth} and point cloud. Extensive experiments demonstrate the effectiveness of the proposed AnA framework on learning plasticity and memory stability during continual learning. Furthermore, PathWeave performs comparably to state-of-the-art MLLMs while concurrently reducing parameter training burdens by 98.73\%. Our code locates at \url{https://github.com/JiazuoYu/PathWeave}.

PAnoramic Semantic Segmentation (PASS) is an important task in computer vision, as it enables semantic understanding of a 360° environment. Currently, most of existing works have focused on addressing the distortion issues in 2D panoramic images without considering spatial properties of indoor scene. This restricts PASS methods in perceiving contextual attributes to deal with the ambiguity when working with monocular images. In this paper, we propose a novel approach for indoor panoramic semantic segmentation. Unlike previous works, we consider the panoramic image as a composition of segment groups: oversampled segments, representing planar structures such as floors and ceilings, and under-sampled segments, representing other scene elements. To optimize each group, we first enhance over-sampled segments by jointly optimizing with a dense depth estimation task. Then, we introduce a transformer-based context module that aggregates different geometric representations of the scene, combined with a simple high-resolution branch, it serves as a robust hybrid decoder for estimating under-sampled segments, effectively preserving the resolution of predicted masks while leveraging various indoor geometric properties. Experimental results on both real-world (Stanford2D3DS, Matterport3D) and synthetic (Structured3D) datasets demonstrate the robustness of our framework, by setting new state-of-the-arts in almost evaluations, The code and updated results are available at: https://github.com/caodinhduc/vertical_relative_distance.

Large-vocabulary object detectors (LVDs) aim to detect objects of many categories, which learn super objectness features and can locate objects accurately while applied to various downstream data. However, LVDs often struggle in recognizing the located objects due to domain discrepancy in data distribution and object vocabulary. At the other end, recent vision-language foundation models such as CLIP demonstrate superior open-vocabulary recognition capability. This paper presents KGD, a Knowledge Graph Distillation technique that exploits the implicit knowledge graphs (KG) in CLIP for effectively adapting LVDs to various downstream domains. KGD consists of two consecutive stages: 1) KG extraction that employs CLIP to encode downstream domain data as nodes and their feature distances as edges, constructing KG that inherits the rich semantic relations in CLIP explicitly; and 2) KG encapsulation that transfers the extracted KG into LVDs to enable accurate cross-domain object classification. In addition, KGD can extract both visual and textual KG independently, providing complementary vision and language knowledge for object localization and object classification in detection tasks over various downstream domains. Experiments over multiple widely adopted detection benchmarks show that KGD outperforms the state-of-the-art consistently by large margins. Codes will be released.

A comprehensive benchmark is yet to be established in the Image Manipulation Detection \& Localization (IMDL) field. The absence of such a benchmark leads to insufficient and misleading model evaluations, severely undermining the development of this field. However, the scarcity of open-sourced baseline models and inconsistent training and evaluation protocols make conducting rigorous experiments and faithful comparisons among IMDL models challenging. To address these challenges, we introduce IMDL-BenCo, the first comprehensive IMDL benchmark and modular codebase. IMDL-BenCo: i) decomposes the IMDL framework into standardized, reusable components and revises the model construction pipeline, improving coding efficiency and customization flexibility; ii) fully implements or incorporates training code for state-of-the-art models to establish a comprehensive IMDL benchmark; and iii) conducts deep analysis based on the established benchmark and codebase, offering new insights into IMDL model architecture, dataset characteristics, and evaluation standards. Specifically, IMDL-BenCo includes common processing algorithms, 8 state-of-the-art IMDL models (1 of which are reproduced from scratch), 2 sets of standard training and evaluation protocols, 15 GPU-accelerated evaluation metrics, and 3 kinds of robustness evaluation. This benchmark and codebase represent a significant leap forward in calibrating the current progress in the IMDL field and inspiring future breakthroughs. Code is available at: https://github.com/scu-zjz/IMDLBenCo