NeurIPS 2024 Papers

Existing open-set learning methods consider only the single-layer labels of objects and strictly assume no overlap between the training and testing sets, leading to contradictory optimization for superposed categories. In this paper, we introduce a more practical Semi-Open Environment setting for open-set 3D object retrieval with hierarchical labels, in which the training and testing set share a partial label space for coarse categories but are completely disjoint from fine categories. We propose the Hypergraph-Based Hierarchical Equilibrium Representation (HERT) framework for this task. Specifically, we propose the Hierarchical Retrace Embedding (HRE) module to overcome the global disequilibrium of unseen categories by fully leveraging the multi-level category information. Besides, tackling the feature overlap and class confusion problem, we perform the Structured Equilibrium Tuning (SET) module to utilize more equilibrial correlations among objects and generalize to unseen categories, by constructing a superposed hypergraph based on the local coherent and global entangled correlations. Furthermore, we generate four semi-open 3DOR datasets with multi-level labels for benchmarking. Results demonstrate that the proposed method can effectively generate the hierarchical embeddings of 3D objects and generalize them towards semi-open environments.

Sketch Re-identification (Sketch Re-ID), which aims to retrieve target person from an image gallery based on a sketch query, is crucial for criminal investigation, law enforcement, and missing person searches. Existing methods aim to alleviate the modality gap by employing semantic metrics constraints or auxiliary modal guidance. However, they incur expensive labor costs and inevitably omit fine-grained modality-consistent information due to the abstraction of sketches. To address this issue, this paper proposes a novel $\textit{Optimal Transport-based Labor-free Text Prompt Modeling}$ (OLTM) network, which hierarchically extracts coarse- and fine-grained similarity representations guided by textual semantic information without any additional annotations. Specifically, multiple target attributes are flexibly obtained by a pre-trained visual question answering (VQA) model. Subsequently, a text prompt reasoning module employs learnable prompt strategy and optimal transport algorithm to extract discriminative global and local text representations, which serve as a bridge for hierarchical and multi-granularity modal alignment between sketch and image modalities. Additionally, instead of measuring the similarity of two samples by only computing their distance, a novel triplet assignment loss is further proposed, in which the whole data distribution also contributes to optimizing the inter/intra-class distances. Extensive experiments conducted on two public benchmarks consistently demonstrate the robustness and superiority of our OLTM over state-of-the-art methods.

The expensive fine-grained annotation and data scarcity have become the primary obstacles for the widespread adoption of deep learning-based Whole Slide Images (WSI) classification algorithms in clinical practice. Unlike few-shot learning methods in natural images that can leverage the labels of each image, existing few-shot WSI classification methods only utilize a small number of fine-grained labels or weakly supervised slide labels for training in order to avoid expensive fine-grained annotation. They lack sufficient mining of available WSIs, severely limiting WSI classification performance. To address the above issues, we propose a novel and efficient dual-tier few-shot learning paradigm for WSI classification, named FAST. FAST consists of a dual-level annotation strategy and a dual-branch classification framework. Firstly, to avoid expensive fine-grained annotation, we collect a very small number of WSIs at the slide level, and annotate an extremely small number of patches. Then, to fully mining the available WSIs, we use all the patches and available patch labels to build a cache branch, which utilizes the labeled patches to learn the labels of unlabeled patches and through knowledge retrieval for patch classification. In addition to the cache branch, we also construct a prior branch that includes learnable prompt vectors, using the text encoder of visual-language models for patch classification. Finally, we integrate the results from both branches to achieve WSI classification. Extensive experiments on binary and multi-class datasets demonstrate that our proposed method significantly surpasses existing few-shot classification methods and approaches the accuracy of fully supervised methods with only 0.22% annotation costs. All codes and models will be publicly available on https://github.com/fukexue/FAST.

Vision-based Semantic Scene Completion (SSC) has gained much attention due to its widespread applications in various 3D perception tasks. Existing sparse-to-dense approaches typically employ shared context-independent queries across various input images, which fails to capture distinctions among them as the focal regions of different inputs vary and may result in undirected feature aggregation of cross-attention. Additionally, the absence of depth information may lead to points projected onto the image plane sharing the same 2D position or similar sampling points in the feature map, resulting in depth ambiguity. In this paper, we present a novel context and geometry aware voxel transformer. It utilizes a context aware query generator to initialize context-dependent queries tailored to individual input images, effectively capturing their unique characteristics and aggregating information within the region of interest. Furthermore, it extend deformable cross-attention from 2D to 3D pixel space, enabling the differentiation of points with similar image coordinates based on their depth coordinates. Building upon this module, we introduce a neural network named CGFormer to achieve semantic scene completion. Simultaneously, CGFormer leverages multiple 3D representations (i.e., voxel and TPV) to boost the semantic and geometric representation abilities of the transformed 3D volume from both local and global perspectives. Experimental results demonstrate that CGFormer achieves state-of-the-art performance on the SemanticKITTI and SSCBench-KITTI-360 benchmarks, attaining a mIoU of 16.87 and 20.05, as well as an IoU of 45.99 and 48.07, respectively. Remarkably, CGFormer even outperforms approaches employing temporal images as inputs or much larger image backbone networks.

Mamba, a state-space model with selective mechanisms and hardware-aware architecture, has demonstrated outstanding performance in long sequence modeling tasks, particularly garnering widespread exploration and application in the field of computer vision. While existing works have mixed opinions of its application in visual tasks, the exploration of its internal workings and the optimization of its performance remain urgent and worthy research questions given its status as a novel model. Existing optimizations of the Mamba model, especially when applied in the visual domain, have primarily relied on predefined methods such as improving scanning mechanisms or integrating other architectures, often requiring strong priors and extensive trial and error. In contrast to these approaches, this paper proposes the Vision Mamba Mender, a systematic approach for understanding the workings of Mamba, identifying flaws within, and subsequently optimizing model performance. Specifically, we present methods for predictive correlation analysis of Mamba's hidden states from both internal and external perspectives, along with corresponding definitions of correlation scores, aimed at understanding the workings of Mamba in visual recognition tasks and identifying flaws therein. Additionally, tailored repair methods are proposed for identified external and internal state flaws to eliminate them and optimize model performance. Extensive experiments validate the efficacy of the proposed methods on prevalent Mamba architectures, significantly enhancing Mamba's performance.

Web-crawled pretraining datasets underlie the impressive "zero-shot" evaluation performance of multimodal models, such as CLIP for classification and Stable-Diffusion for image generation. However, it is unclear how meaningful the notion of "zero-shot" generalization is for such multimodal models, as it is not known to what extent their pretraining datasets encompass the downstream concepts targeted for during "zero-shot" evaluation. In this work, we ask: How is the performance of multimodal models on downstream concepts influenced by the frequency of these concepts in their pretraining datasets? We comprehensively investigate this question across 34 models and 5 standard pretraining datasets (CC-3M, CC-12M, YFCC-15M, LAION-400M, LAION-Aesthetics), generating over 300GB of data artifacts. We consistently find that, far from exhibiting "zero-shot" generalization, multimodal models require exponentially more data to achieve linear improvements in downstream "zero-shot" performance, following a sample inefficient log-linear scaling trend. This trend persists even when controlling for sample-level similarity between pretraining and downstream datasets, and testing on purely synthetic data distributions. Furthermore, upon benchmarking models on long-tailed data sampled based on our analysis, we demonstrate that multimodal models across the board perform poorly. We contribute this long-tail test set as the Let it Wag! benchmark to further research in this direction. Taken together, our study reveals an exponential need for training data which implies that the key to "zero-shot" generalization capabilities under large-scale training data and compute paradigms remains to be found.

Although Vision Transformers (ViTs) have recently advanced computer vision tasks significantly, an important real-world problem was overlooked: adapting to variable input resolutions. Typically, images are resized to a fixed resolution, such as 224x224, for efficiency during training and inference. However, uniform input size conflicts with real-world scenarios where images naturally vary in resolution. Modifying the preset resolution of a model may severely degrade the performance. In this work, we propose to enhance the model adaptability to resolution variation by optimizing the patch embedding. The proposed method, called Multi-Scale Patch Embedding (MSPE), substitutes the standard patch embedding with multiple variable-sized patch kernels and selects the best parameters for different resolutions, eliminating the need to resize the original image. Our method does not require high-cost training or modifications to other parts, making it easy to apply to most ViT models. Experiments in image classification, segmentation, and detection tasks demonstrate the effectiveness of MSPE, yielding superior performance on low-resolution inputs and performing comparably on high-resolution inputs with existing methods.

Open-World Object Detection (OWOD) is a challenging task that requires the detector to identify unlabeled objects and continuously demands the detector to learn new knowledge based on existing ones. Existing methods primarily focus on recalling unknown objects, neglecting to explore the reasons behind them. This paper aims to understand the model's behavior in predicting the unknown category. First, we model the text attribute and the positive sample probability, obtaining their empirical probability, which can be seen as the detector's estimation of the likelihood of the target with certain known attributes being predicted as the foreground. Then, we jointly decide whether the current object should be categorized in the unknown category based on the empirical, the in-distribution, and the out-of-distribution probability. Finally, based on the decision-making process, we can infer the similarity of an unknown object to known classes and identify the attribute with the most significant impact on the decision-making process. This additional information can help us understand the behavior of the model's prediction in the unknown class. The evaluation results on the Real-World Object Detection (RWD) benchmark, which consists of five real-world application datasets, show that we surpassed the previous state-of-the-art (SOTA) with an absolute gain of 5.3 mAP for unknown classes, reaching 20.5 mAP. Our code is available at https://github.com/xxyzll/UMB.

Promptable segmentation typically requires instance-specific manual prompts to guide the segmentation of each desired object. To minimize such a need, task-generic promptable segmentation has been introduced, which employs a single task-generic prompt to segment various images of different objects in the same task. Current methods use Multimodal Large Language Models (MLLMs) to reason detailed instance-specific prompts from a task-generic prompt for improving segmentation accuracy. The effectiveness of this segmentation heavily depends on the precision of these derived prompts. However, MLLMs often suffer hallucinations during reasoning, resulting in inaccurate prompting. While existing methods focus on eliminating hallucinations to improve a model, we argue that MLLM hallucinations can reveal valuable contextual insights when leveraged correctly, as they represent pre-trained large-scale knowledge beyond individual images. In this paper, we first utilize hallucinations to mine task-related information from images and verify its accuracy to enhance precision of the generated prompts. Specifically, we introduce an iterative \textbf{Pro}mpt-\textbf{Ma}sk \textbf{C}ycle generation framework (ProMaC) with a prompt generator and a mask generator. The prompt generator uses a multi-scale chain of thought prompting, initially leveraging hallucinations to extract extended contextual prompts on a test image. These hallucinations are then minimized to formulate precise instance-specific prompts, directing the mask generator to produce masks that are consistent with task semantics by mask semantic alignment. Iteratively the generated masks induce the prompt generator to focus more on task-relevant image areas and reduce irrelevant hallucinations, resulting jointly in better prompts and masks. Experiments on 5 benchmarks demonstrate the effectiveness of ProMaC. Code is in https://lwpyh.github.io/ProMaC/.

Reconstructing a sequence of sharp images from the blurry input is crucial for enhancing our insights into the captured scene and poses a significant challenge due to the limited temporal features embedded in the image. Spike cameras, sampling at rates up to 40,000 Hz, have proven effective in capturing motion features and beneficial for solving this ill-posed problem. Nonetheless, existing methods fall into the supervised learning paradigm, which suffers from notable performance degradation when applied to real-world scenarios that diverge from the synthetic training data domain. To address these challenges, we propose the first self-supervised framework for the task of spike-guided motion deblurring. Our approach begins with the formulation of a spike-guided deblurring model that explores the theoretical relationships among spike streams, blurry images, and their corresponding sharp sequences. We subsequently develop a self-supervised cascaded framework to alleviate the issues of spike noise and spatial-resolution mismatching encountered in the deblurring model. With knowledge distillation and re-blurring loss, we further design a lightweight deblur network to generate high-quality sequences with brightness and texture consistency with the original input. Quantitative and qualitative experiments conducted on our real-world and synthetic datasets with spikes validate the superior generalization of the proposed framework. Our code, data and trained models are available at \url{https://github.com/chenkang455/S-SDM}.

We propose a novel unsupervised method to learn pose and part-segmentation of articulated objects with rigid parts. Given two observations of an object in different articulation states, our method learns the geometry and appearance of object parts by using an implicit model from the first observation, distills the part segmentation and articulation from the second observation while rendering the latter observation. Additionally, to tackle the complexities in the joint optimization of part segmentation and articulation, we propose a voxel grid based initialization strategy and a decoupled optimization procedure. Compared to the prior unsupervised work, our model obtains significantly better performance, generalizes to objects with multiple parts while it can be efficiently from few views for the latter observation.

Large Language Models (LLMs) have demonstrated impressive performance on multimodal tasks, without any multimodal finetuning. They are the de facto building block for Large Multimodal Models (LMMs), yet, we still lack a proper understanding of their success. In this work, we expose frozen LLMs to image, video, audio and text inputs and analyse their internal representation with the attempt to understand their generalization beyond textual inputs. Our work provides the following **findings.** Perceptual tokens (1) are easily distinguishable from textual ones inside LLMs, with significantly different representations (e.g. live in different narrow cones), and complete translation to textual tokens does not exists. Yet, (2) both perceptual and textual tokens activate similar LLM weights. Despite their differences, (3) perceptual tokens are implicitly aligned to textual tokens inside LLMs, we call this the implicit multimodal alignment effect (IMA), and argue that this is linked to architectural design, helping LLMs to generalize. This provide more evidence to believe that the generalization of LLMs to multimodal inputs is mainly due to their architecture. These findings lead to several **implications.** This work provides several implications. (1) We find a positive correlation between the implicit alignment score and the task performance, suggesting that this could act as a proxy metric for model evaluation and selection. (2) A negative correlation exists regarding hallucinations (e.g. describing non-existing objects in images), revealing that this problem is mainly due to misalignment between the internal perceptual and textual representations. (3) Perceptual tokens change slightly throughout the model, thus, we propose different approaches to skip computations (e.g. in FFN layers), and significantly reduce the inference cost. (4) Due to the slowly changing embeddings across layers, and the high overlap between textual and multimodal activated weights, we compress LLMs by keeping only 1 subnetwork (called alpha-SubNet) that works well across a wide range of multimodal tasks. The code is available here: https://github.com/mshukor/ima-lmms.

Unsupervised 3D object detection methods have emerged to leverage vast amounts of data without requiring manual labels for training. Recent approaches rely on dynamic objects for learning to detect mobile objects but penalize the detections of static instances during training. Multiple rounds of (self) training are used to add detected static instances to the set of training targets; this procedure to improve performance is computationally expensive. To address this, we propose the method UNION. We use spatial clustering and self-supervised scene flow to obtain a set of static and dynamic object proposals from LiDAR. Subsequently, object proposals' visual appearances are encoded to distinguish static objects in the foreground and background by selecting static instances that are visually similar to dynamic objects. As a result, static and dynamic mobile objects are obtained together, and existing detectors can be trained with a single training. In addition, we extend 3D object discovery to detection by using object appearance-based cluster labels as pseudo-class labels for training object classification. We conduct extensive experiments on the nuScenes dataset and increase the state-of-the-art performance for unsupervised 3D object discovery, i.e. UNION more than doubles the average precision to 38.4. The code is available at github.com/TedLentsch/UNION.

Adaptation of pretrained vision-language models such as CLIP to various downstream tasks have raised great interest in recent researches. Previous works have proposed a variety of test-time adaptation (TTA) methods to achieve strong generalization without any knowledge of the target domain. However, existing training-required TTA approaches like TPT necessitate entropy minimization that involves large computational overhead, while training-free methods like TDA overlook the potential for information mining from the test samples themselves. In this paper, we break down the design of existing popular training-required and training-free TTA methods and bridge the gap between them within our framework. Specifically, we maintain a light-weight key-value memory for feature retrieval from instance-agnostic historical samples and instance-aware boosting samples. The historical samples are filtered from the testing data stream and serve to extract useful information from the target distribution, while the boosting samples are drawn from regional bootstrapping and capture the knowledge of the test sample itself. We theoretically justify the rationality behind our method and empirically verify its effectiveness on both the out-of-distribution and the cross-domain datasets, showcasing its applicability in real-world situations.

Existing prompt-tuning methods have demonstrated impressive performances in continual learning (CL), by selecting and updating relevant prompts in the vision-transformer models. On the contrary, this paper aims to learn each task by tuning the prompts in the direction orthogonal to the subspace spanned by previous tasks' features, so as to ensure no interference on tasks that have been learned to overcome catastrophic forgetting in CL. However, different from the orthogonal projection in the traditional CNN architecture, the prompt gradient orthogonal projection in the ViT architecture shows completely different and greater challenges, i.e., 1) the high-order and non-linear self-attention operation; 2) the drift of prompt distribution brought by the LayerNorm in the transformer block. Theoretically, we have finally deduced two consistency conditions to achieve the prompt gradient orthogonal projection, which provide a theoretical guarantee of eliminating interference on previously learned knowledge via the self-attention mechanism in visual prompt tuning. In practice, an effective null-space-based approximation solution has been proposed to implement the prompt gradient orthogonal projection. Extensive experimental results demonstrate the effectiveness of anti-forgetting on four class-incremental benchmarks with diverse pre-trained baseline models, and our approach achieves superior performances to state-of-the-art methods. Our code is available at https://github.com/zugexiaodui/VPTinNSforCL

The modern study and use of surfaces is a research topic grounded in centuries of mathematical and empirical inquiry. From a mathematical point of view, curvature is an invariant that characterises the intrinsic geometry and the extrinsic shape of a surface. Yet, in modern applications the focus has shifted away from finding expressive representations of surfaces, and towards the design of efficient neural network architectures to process them. The literature suggests a tendency to either overlook the representation of the processed surface, or use overcomplicated representations whose ability to capture the essential features of a surface is opaque. We propose using curvature as the input of neural network architectures for surface processing, and explore this proposition through experiments making use of the shape operator. Our results show that using curvature as input leads to significant a increase in performance on segmentation and classification tasks, while allowing far less computational overhead than current methods.

Recent advancements in anomaly detection have seen the efficacy of CNN- and transformer-based approaches. However, CNNs struggle with long-range dependencies, while transformers are burdened by quadratic computational complexity. Mamba-based models, with their superior long-range modeling and linear efficiency, have garnered substantial attention. This study pioneers the application of Mamba to multi-class unsupervised anomaly detection, presenting MambaAD, which consists of a pre-trained encoder and a Mamba decoder featuring (Locality-Enhanced State Space) LSS modules at multi-scales. The proposed LSS module, integrating parallel cascaded (Hybrid State Space) HSS blocks and multi-kernel convolutions operations, effectively captures both long-range and local information. The HSS block, utilizing (Hybrid Scanning) HS encoders, encodes feature maps into five scanning methods and eight directions, thereby strengthening global connections through the (State Space Model) SSM. The use of Hilbert scanning and eight directions significantly improves feature sequence modeling. Comprehensive experiments on six diverse anomaly detection datasets and seven metrics demonstrate state-of-the-art performance, substantiating the method's effectiveness. The code and models are available at https://lewandofskee.github.io/projects/MambaAD.

We introduce ReplaceAnything3D model RAM3D, a novel method for 3D object replacement in 3D scenes based on users' text description. Given multi-view images of a scene, a text prompt describing the object to replace, and another describing the new object, our Erase-and-Replace approach can effectively swap objects in 3D scenes with newly generated content while maintaining 3D consistency across multiple viewpoints. We demonstrate the versatility of RAM3D by applying it to various realistic 3D scene types, showcasing results of modified objects that blend in seamlessly with the scene without impacting its overall integrity.