NeurIPS 2024 Papers

We propose a novel and general method to learn Bregman divergences from raw high-dimensional data that measure similarity between images in pixel space. As a prototypical application, we learn divergences that consider real-world corruptions of images (e.g., blur) as close to the original and noisy perturbations as far, even if in $L^p$-distance the opposite holds. We also show that the learned Bregman divergence excels on datasets of human perceptual similarity judgment, suggesting its utility in a range of applications. We then define adversarial attacks by replacing the projected gradient descent (PGD) with the mirror descent associated with the learned Bregman divergence, and use them to improve the state-of-the-art in robustness through adversarial training for common image corruptions. In particular, for the contrast corruption that was found problematic in prior work we achieve an accuracy that exceeds the $L^p$- and the LPIPS-based adversarially trained neural networks by a margin of 27.16\% on the CIFAR-10-C corruption data set.

Atmospheric turbulence, caused by random fluctuations in the atmosphere's refractive index, introduces complex spatio-temporal distortions in imagery captured at long range. Video Atmospheric Turbulence Mitigation (ATM) aims to restore videos affected by these distortions. However, existing video ATM methods, both supervised and self-supervised, struggle to maintain temporally consistent mitigation across frames, leading to visually incoherent results. This limitation arises from the stochastic nature of atmospheric turbulence, which varies across space and time. Inspired by the observation that atmospheric turbulence induces high-frequency temporal variations, we propose ConVRT, a novel framework for consistent video restoration through turbulence. ConVRT introduces a neural video representation that explicitly decouples spatial and temporal information into a spatial content field and a temporal deformation field, enabling targeted regularization of the network's temporal representation capability. By leveraging the low-pass filtering properties of the regularized temporal representations, ConVRT effectively mitigates turbulence-induced temporal frequency variations and promotes temporal consistency. Furthermore, our training framework seamlessly integrates supervised pre-training on synthetic turbulence data with self-supervised learning on real-world videos, significantly improving the temporally consistent mitigation of ATM methods on diverse real-world data. More information can be found on our project page: https://convrt-2024.github.io/

Incorporating inductive bias by embedding geometric entities (such as rays) as input has proven successful in multi-view learning. However, the methods adopting this technique typically lack equivariance, which is crucial for effective 3D learning. Equivariance serves as a valuable inductive prior, aiding in the generation of robust multi-view features for 3D scene understanding. In this paper, we explore the application of equivariant multi-view learning to depth estimation, not only recognizing its significance for computer vision and robotics but also addressing the limitations of previous research. Most prior studies have either overlooked equivariance in this setting or achieved only approximate equivariance through data augmentation, which often leads to inconsistencies across different reference frames. To address this issue, we propose to embed $SE(3)$ equivariance into the Perceiver IO architecture. We employ Spherical Harmonics for positional encoding to ensure 3D rotation equivariance, and develop a specialized equivariant encoder and decoder within the Perceiver IO architecture. To validate our model, we applied it to the task of stereo depth estimation, achieving state of the art results on real-world datasets without explicit geometric constraints or extensive data augmentation.

Given different instructions, large vision-language models (LVLMs) exhibit different degrees of object hallucinations, posing a significant challenge to the evaluation of object hallucinations. Overcoming this challenge, existing object hallucination evaluation methods average the results obtained from a set of instructions. However, these methods fail to provide consistent evaluation across instruction sets that generate image descriptions of significantly different lengths. In this paper, we present the first systematic investigation of the effect of instructions on object hallucinations in LVLMs, with a specific focus on the role played by image description lengths. A valuable finding is that instructions indirectly affect hallucinations through the length of image descriptions. The longer the image description, the higher the object hallucination degree. Accordingly, we fit an informative length-hallucination curve, upon which a fine-grained evaluation framework named LeHaCE is introduced for evaluating object hallucinations at any given image description length. LeHaCE evaluates the object hallucination degree at a uniform image description length to mitigate the effect of description lengths, promoting stability and fairness. Moreover, LeHaCE incorporates the curve slope as an innovative hallucination evaluation metric, reflecting the extent to which the object hallucination degree is affected by the image description length, achieving a more comprehensive evaluation. Experimental results demonstrate that LeHaCE provides a more stable, fair, and comprehensive evaluation of object hallucinations in LVLMs compared to existing methods.

Existing multi-modal image fusion methods fail to address the compound degradations presented in source images, resulting in fusion images plagued by noise, color bias, improper exposure, etc. Additionally, these methods often overlook the specificity of foreground objects, weakening the salience of the objects of interest within the fused images. To address these challenges, this study proposes a novel interactive multi-modal image fusion framework based on the text-modulated diffusion model, called Text-DiFuse. First, this framework integrates feature-level information integration into the diffusion process, allowing adaptive degradation removal and multi-modal information fusion. This is the first attempt to deeply and explicitly embed information fusion within the diffusion process, effectively addressing compound degradation in image fusion. Second, by embedding the combination of the text and zero-shot location model into the diffusion fusion process, a text-controlled fusion re-modulation strategy is developed. This enables user-customized text control to improve fusion performance and highlight foreground objects in the fused images. Extensive experiments on diverse public datasets show that our Text-DiFuse achieves state-of-the-art fusion performance across various scenarios with complex degradation. Moreover, the semantic segmentation experiment validates the significant enhancement in semantic performance achieved by our text-controlled fusion re-modulation strategy. The code is publicly available at https://github.com/Leiii-Cao/Text-DiFuse.

Diffusion Transformer (DiT) is emerging as a cutting-edge trend in the landscape of generative diffusion models for image generation. Recently, masked-reconstruction strategies have been considered to improve the efficiency and semantic consistency in training DiT but suffer from deficiency in contextual information extraction. In this paper, we provide a new insight to reveal that noisy-to-noisy masked-reconstruction harms sufficient utilization of contextual information. We further demonstrate the insight with theoretical analysis and empirical study on the mutual information between unmasked and masked patches. Guided by such insight, we propose a novel training paradigm named MC-DiT for fully learning contextual information via diffusion denoising at different noise variances with clean-to-clean mask-reconstruction. Moreover, to avoid model collapse, we design two complementary branches of DiT decoders for enhancing the use of noisy patches and mitigating excessive reliance on clean patches in reconstruction. Extensive experimental results on 256$\times$256 and 512$\times$512 image generation on the ImageNet dataset demonstrate that the proposed MC-DiT achieves state-of-the-art performance in unconditional and conditional image generation with enhanced convergence speed.

Implicit representation has opened up new possibilities for inverse rendering. However, existing implicit neural inverse rendering methods struggle to handle strongly illuminated scenes with significant shadows and slight reflections. The existence of shadows and reflections can lead to an inaccurate understanding of the scene, making precise factorization difficult. To this end, we present RobIR, an implicit inverse rendering approach that uses ACES tone mapping and regularized visibility estimation to reconstruct accurate BRDF of the object. By accurately modeling the indirect radiance field, normal, visibility, and direct light simultaneously, we are able to accurately decouple environment lighting and the object's PBR materials without imposing strict constraints on the scene. Even in high-illumination scenes with shadows and specular reflections, our method can recover high-quality albedo and roughness with no shadow interference. RobIR outperforms existing methods in both quantitative and qualitative evaluations.

Anchoring is a recent, architecture-agnostic principle for training deep neural networks that has been shown to significantly improve uncertainty estimation, calibration, and extrapolation capabilities. In this paper, we systematically explore anchoring as a general protocol for training vision models, providing fundamental insights into its training and inference processes and their implications for generalization and safety. Despite its promise, we identify a critical problem in anchored training that can lead to an increased risk of learning undesirable shortcuts, thereby limiting its generalization capabilities. To address this, we introduce a new anchored training protocol that employs a simple regularizer to mitigate this issue and significantly enhances generalization. We empirically evaluate our proposed approach across datasets and architectures of varying scales and complexities, demonstrating substantial performance gains in generalization and safety metrics compared to the standard training protocol. The open-source code is available at https://software.llnl.gov/anchoring.

We propose ManiPose, a manifold-constrained multi-hypothesis model for human-pose 2D-to-3D lifting. We provide theoretical and empirical evidence that, due to the depth ambiguity inherent to monocular 3D human pose estimation, traditional regression models suffer from pose-topology consistency issues, which standard evaluation metrics (MPJPE, P-MPJPE and PCK) fail to assess. ManiPose addresses depth ambiguity by proposing multiple candidate 3D poses for each 2D input, each with its estimated plausibility. Unlike previous multi-hypothesis approaches, ManiPose forgoes generative models, greatly facilitating its training and usage. By constraining the outputs to lie on the human pose manifold, ManiPose guarantees the consistency of all hypothetical poses, in contrast to previous works. We showcase the performance of ManiPose on real-world datasets, where it outperforms state-of-the-art models in pose consistency by a large margin while being very competitive on the MPJPE metric.

Collaborative perception is dedicated to tackling the constraints of single-agent perception, such as occlusions, based on the multiple agents' multi-view sensor inputs. However, most existing works assume an ideal condition that all agents' multi-view cameras are continuously available. In reality, cameras may be highly noisy, obscured or even failed during the collaboration. In this work, we introduce a new robust camera-insensitivity problem: how to overcome the issues caused by the failed camera perspectives, while stabilizing high collaborative performance with low calibration cost? To address above problems, we propose RCDN, a Robust Camera-insensitivity collaborative perception with a novel Dynamic feature-based 3D Neural modeling mechanism. The key intuition of RCDN is to construct collaborative neural rendering field representations to recover failed perceptual messages sent by multiple agents. To better model collaborative neural rendering field, RCDN first establishes a geometry BEV feature based time-invariant static field with other agents via fast hash grid modeling. Based on the static background field, the proposed time-varying dynamic field can model corresponding motion vector for foregrounds with appropriate positions. To validate RCDN, we create OPV2V-N, a new large-scale dataset with manual labelling under different camera failed scenarios. Extensive experiments conducted on OPV2V-N show that RCDN can be ported to other baselines and improve their robustness in extreme camera-insensitivity setting. Our code and datasets will be available soon.

This paper addresses the problem of on-road object importance estimation, which utilizes video sequences captured from the driver's perspective as the input. Although this problem is significant for safer and smarter driving systems, the exploration of this problem remains limited. On one hand, publicly-available large-scale datasets are scarce in the community. To address this dilemma, this paper contributes a new large-scale dataset named Traffic Object Importance (TOI). On the other hand, existing methods often only consider either bottom-up feature or single-fold guidance, leading to limitations in handling highly dynamic and diverse traffic scenarios. Different from existing methods, this paper proposes a model that integrates multi-fold top-down guidance with the bottom-up feature. Specifically, three kinds of top-down guidance factors (i.e., driver intention, semantic context, and traffic rule) are integrated into our model. These factors are important for object importance estimation, but none of the existing methods simultaneously consider them. To our knowledge, this paper proposes the first on-road object importance estimation model that fuses multi-fold top-down guidance factors with bottom-up feature. Extensive experiments demonstrate that our model outperforms state-of-the-art methods by large margins, achieving 23.1% Average Precision (AP) improvement compared with the recently proposed model (i.e., Goal).

Knowledge Distillation is the mechanism by which the insights gained from a larger teacher model are transferred to a smaller student model. However, the transfer suffers when the teacher model is significantly larger than the student. To overcome this, prior works have proposed training intermediately sized models, Teacher Assistants (TAs) to help the transfer process. However, training TAs is expensive, as training these models is a knowledge transfer task in itself. Further, these TAs are larger than the student model and training them especially in large data settings can be computationally intensive. In this paper, we propose a novel framework called Controlled Information Flow for Knowledge Distillation (CIFD) consisting of two components. First, we propose a significantly smaller alternatives to TAs, the Rate-Distortion Module (RDM) which uses the teacher's penultimate layer embedding and a information rate-constrained bottleneck layer to replace the Teacher Assistant model. RDMs are smaller and easier to train than TAs, especially in large data regimes, since they operate on the teacher embeddings and do not need to relearn low level input feature extractors. Also, by varying the information rate across the bottleneck, RDMs can replace TAs of different sizes. Secondly, we propose the use of Information Bottleneck Module in the student model, which is crucial for regularization in the presence of a large number of RDMs. We show comprehensive state-of-the-art results of the proposed method over large datasets like Imagenet. Further, we show the significant improvement in distilling CLIP like models over a huge 12M image-text dataset. It outperforms CLIP specialized distillation methods across five zero-shot classification datasets and two zero-shot image-text retrieval datasets.

In the field of 3D Human Pose Estimation (HPE), scalability and generalization across diverse real-world scenarios remain significant challenges. This paper addresses two key bottlenecks to scalability: limited data diversity caused by 'popularity bias' and increased 'one-to-many' depth ambiguity arising from greater pose diversity. We introduce the Biomechanical Pose Generator (BPG), which leverages biomechanical principles, specifically the normal range of motion, to autonomously generate a wide array of plausible 3D poses without relying on a source dataset, thus overcoming the restrictions of popularity bias. To address depth ambiguity, we propose the Binary Depth Coordinates (BDC), which simplifies depth estimation into a binary classification of joint positions (front or back). This method decomposes a 3D pose into three core elements—2D pose, bone length, and binary depth decision—substantially reducing depth ambiguity and enhancing model robustness and accuracy, particularly in complex poses. Our results demonstrate that these approaches increase the diversity and volume of pose data while consistently achieving performance gains, even amid the complexities introduced by increased pose diversity.

Current approaches for open-vocabulary scene graph generation (OVSGG) use vision-language models such as CLIP and follow a standard zero-shot pipeline – computing similarity between the query image and the text embeddings for each category (i.e., text classifiers). In this work, we argue that the text classifiers adopted by existing OVSGG methods, i.e., category-/part-level prompts, are scene-agnostic as they remain unchanged across contexts. Using such fixed text classifiers not only struggles to model visual relations with high variance, but also falls short in adapting to distinct contexts. To plug these intrinsic shortcomings, we devise SDSGG, a scene-specific description based OVSGG framework where the weights of text classifiers are adaptively adjusted according to the visual content. In particular, to generate comprehensive and diverse descriptions oriented to the scene, an LLM is asked to play different roles (e.g., biologist and engineer) to analyze and discuss the descriptive features of a given scene from different views. Unlike previous efforts simply treating the generated descriptions as mutually equivalent text classifiers, SDSGG is equipped with an advanced renormalization mechanism to adjust the influence of each text classifier based on its relevance to the presented scene (this is what the term “specific” means). Furthermore, to capture the complicated interplay between subjects and objects, we propose a new lightweight module called mutual visual adapter. It refines CLIP’s ability to recognize relations by learning an interaction-aware semantic space. Extensive experiments on prevalent benchmarks show that SDSGG significantly outperforms top-leading methods.

Visual-Language Alignment (VLA) has gained a lot of attention since CLIP's groundbreaking work. Although CLIP performs well, the typical direct latent feature alignment lacks clarity in its representation and similarity scores. On the other hand, lexical representation, a vector whose element represents the similarity between the sample and a word from the vocabulary, is a natural sparse representation and interpretable, providing exact matches for individual words. However, lexical representations are difficult to learn due to no ground-truth supervision and false-discovery issues, and thus requires complex design to train effectively. In this paper, we introduce LexVLA, a more interpretable VLA framework by learning a unified lexical representation for both modalities without complex design. We use DINOv2 as our visual model for its local-inclined features and Llama 2, a generative language model, to leverage its in-context lexical prediction ability. To avoid the false discovery, we propose an overuse penalty to refrain the lexical representation from falsely frequently activating meaningless words. We demonstrate that these two pre-trained uni-modal models can be well-aligned by fine-tuning on the modest multi-modal dataset and avoid intricate training configurations. On cross-modal retrieval benchmarks, LexVLA, trained on the CC-12M multi-modal dataset, outperforms baselines fine-tuned on larger datasets (e.g., YFCC15M) and those trained from scratch on even bigger datasets (e.g., 1.1B data, including CC-12M). We conduct extensive experiments to analyze LexVLA. Codes are available at https://github.com/Clementine24/LexVLA.

Natural images captured by mobile devices often suffer from multiple types of degradation, such as noise, blur, and low light. Traditional image restoration methods require manual selection of specific tasks, algorithms, and execution sequences, which is time-consuming and may yield suboptimal results. All-in-one models, though capable of handling multiple tasks, typically support only a limited range and often produce overly smooth, low-fidelity outcomes due to their broad data distribution fitting. To address these challenges, we first define a new pipeline for restoring images with multiple degradations, and then introduce RestoreAgent, an intelligent image restoration system leveraging multimodal large language models. RestoreAgent autonomously assesses the type and extent of degradation in input images and performs restoration through (1) determining the appropriate restoration tasks, (2) optimizing the task sequence, (3) selecting the most suitable models, and (4) executing the restoration. Experimental results demonstrate the superior performance of RestoreAgent in handling complex degradation, surpassing human experts. Furthermore, the system’s modular design facilitates the fast integration of new tasks and models.

We measure the out-of-domain uncertainty in the prediction of Neural Networks using a statistical notion called "Lens Depth'' (LD) combined with Fermat Distance, which is able to capture precisely the "depth'' of a point with respect to a distribution in feature space, without any distributional assumption. Our method also has no trainable parameter. The method is applied directly in the feature space at test time and does not intervene in training process. As such, it does not impact the performance of the original model. The proposed method gives excellent qualitative results on toy datasets and can give competitive or better uncertainty estimation on standard deep learning datasets compared to strong baseline methods.

Social relation reasoning aims to identify relation categories such as friends, spouses, and colleagues from images. While current methods adopt the paradigm of training a dedicated network end-to-end using labeled image data, they are limited in terms of generalizability and interpretability. To address these issues, we first present a simple yet well-crafted framework named SocialGPT, which combines the perception capability of Vision Foundation Models (VFMs) and the reasoning capability of Large Language Models (LLMs) within a modular framework, providing a strong baseline for social relation recognition. Specifically, we instruct VFMs to translate image content into a textual social story, and then utilize LLMs for text-based reasoning. SocialGPT introduces systematic design principles to adapt VFMs and LLMs separately and bridge their gaps. Without additional model training, it achieves competitive zero-shot results on two databases while offering interpretable answers, as LLMs can generate language-based explanations for the decisions. The manual prompt design process for LLMs at the reasoning phase is tedious and an automated prompt optimization method is desired. As we essentially convert a visual classification task into a generative task of LLMs, automatic prompt optimization encounters a unique long prompt optimization issue. To address this issue, we further propose the Greedy Segment Prompt Optimization (GSPO), which performs a greedy search by utilizing gradient information at the segment level. Experimental results show that GSPO significantly improves performance, and our method also generalizes to different image styles. The code is available at https://github.com/Mengzibin/SocialGPT.

Multiple clustering aims to discover various latent structures of data from different aspects. Deep multiple clustering methods have achieved remarkable performance by exploiting complex patterns and relationships in data. However, existing works struggle to flexibly adapt to diverse user-specific needs in data grouping, which may require manual understanding of each clustering. To address these limitations, we introduce Multi-Sub, a novel end-to-end multiple clustering approach that incorporates a multi-modal subspace proxy learning framework in this work. Utilizing the synergistic capabilities of CLIP and GPT-4, Multi-Sub aligns textual prompts expressing user preferences with their corresponding visual representations. This is achieved by automatically generating proxy words from large language models that act as subspace bases, thus allowing for the customized representation of data in terms specific to the user’s interests. Our method consistently outperforms existing baselines across a broad set of datasets in visual multiple clustering tasks. Our code is available at https://github.com/Alexander-Yao/Multi-Sub.