ICCV 2025 论文列表

Foundation medical segmentation models, with MedSAM being the most popular, have achieved promising performance across organs and lesions. However, MedSAM still suffers from compromised performance on specific lesions with intricate structures and appearance, as well as bounding box prompt-induced perturbations. Although current test-time adaptation (TTA) methods for medical image segmentation may tackle this issue, partial (e.g., batch normalization) or whole parametric updates restrict their effectiveness due to limited update signals or catastrophic forgetting in large models. Meanwhile, these approaches ignore the computational complexity during adaptation, which is particularly significant for modern foundation models. To this end, our theoretical analyses reveal that directly refining image embeddings is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance without the risk of catastrophic forgetting. Under this framework, we propose to encourage maximizing factorized conditional probabilities of the posterior prediction probability using a proposed distribution-approximated latent conditional random field loss combined with an entropy minimization loss. Experiments show that we achieve about 3% Dice score improvements across three datasets while reducing computational complexity by over 7 times.

Vision-language models (VLMs) have shown promise in test-time adaptation tasks due to their remarkable capabilities in understanding and reasoning about visual content through natural language descriptions. However, training VLMs typically demands substantial computational resources, and they often struggle to adapt efficiently to new domains or tasks. Additionally, dynamically estimating the test distribution from streaming data at test time remains a significant challenge. In this work, we propose a novel test-time retrieval-augmented adaptation (TT-RAA) method that enables VLMs to maintain high performance across diverse visual recognition tasks without the need for task-specific training or large computational overhead. During inference, TT-RAA employs a streaming mixture of Gaussian database (SMGD) to continuously estimate test distributions, requiring minimal storage. Then, TT-RAA retrieves the most relevant information from the SMGD, enhancing the original VLM outputs. A key limitation of CLIP-based VLMs is their inter-modal vision-language optimization, which does not optimize vision-space similarity, leading to larger intra-modal variance. To address this, we propose a multimodal retrieval augmentation module that transforms the SMGD into a unified multimodal space, enabling retrieval that aligns both vision and language modalities. Extensive experiments across both cross-domain and out-of-distribution benchmarks comprising fourteen datasets demonstrate TT-RAA's superior performance compared to state-of-the-art methods. Ablation studies and hyperparameter analyses further validate the effectiveness of the proposed modules. The source code of our work is available at https://github.com/xinqi-fan/TT-RAA.

Zero-shot depth completion with metric scales poses significant challenges, primarily due to performance limitations such as domain specificity and sensor characteristics. One recent emerging solution is to integrate monocular depth foundation models into depth completion frameworks, yet these efforts still face issues with suboptimal performance and often require further adaptation to the target task. Surprisingly, we find that a simple test-time training, which fine-tunes monocular depth foundation models on sparse depth measurements from sensors just as it is, yields reasonable results. However, this test-time training obviously incurs high computational costs and introduces biases towards specific conditions, making it impractical for real-world scenarios. In this paper, we introduce a new approach toward parameter-efficient zero-shot depth completion. Our key idea of this work is to leverage visual prompt tuning, achieving sensor-specific depth scale adaptation without forgetting foundational knowledge. Experimental results on diverse datasets demonstrate that our approach outperforms relevant state-of-the-art methods, showing superior generalization and efficiency. Our source code is available in the supplementary materials.

We present TerraMind, the first any-to-any generative, multi-modal foundation model for Earth observation (EO). Unlike other multimodal models, TerraMind is pretrained on dual-scale representations combining both token-level and pixel-level data across modalities. On a token level, TerraMind encodes high-level contextual information to learn cross-modal relationships, while on a pixel level, TerraMind leverages fine-grained representations to capture critical spatial nuances. We pretrained TerraMind on nine geospatial modalities of a global, large-scale dataset. In this paper, we demonstrate that (i) TerraMind's dual-scale early fusion approach unlocks a range of zero-shot and few-shot applications for Earth observation, (ii) TerraMind introduces "thinking in modalities" (TiM)--the capability of generating additional artificial data during finetuning and inference to improve the model output--and (iii) TerraMind achieves beyond state-of-the-art performance in community-standard benchmarks for EO like PANGAEA. All models and code have been open-sourced under a permissive license at https://huggingface.co/ibm-esa-geospatial and https://github.com/ibm/terramind.

The combination of Large Language Models (LLMs) and Federated Learning (FL) to leverage privacy-preserving data has emerged as a promising approach to further enhance the Parameter-Efficient Fine-Tuning (PEFT) capabilities of LLMs. In real-world FL settings with resource heterogeneity, the training process of Low-Rank Adaptation (LoRA), the representative PEFT method, still faces two major challenges: aggregation noise and aggregation misalignment. In this paper, we propose a novel Tensor-aggregated LoRA (Te-LoRA) in Federated Fine-tuning based on an alternating-freeze training strategy to avoid aggregating noise without additional server-side computational costs, while mitigating aggregation suboptimality caused by parameter misalignment between heterogeneous LoRAs. Especially in addressing the aggregation suboptimality issue, we design the Pre-Aggregation Alignment strategy (PAA-strategy) and Tensor-to-Matrix strategy (T2M-strategy) for aligning heterogeneous LoRAs and aggregating them into an united tensor, which is then decomposed into matrices adapted for client download. Extensive experiments demonstrate the effectiveness and robustness of Te-LoRA in both homogeneous and heterogeneous settings.

With the rise of applications such as embodied intelligence, developing high real-time online video instance segmentation (VIS) has become increasingly important. However, through time profiling of the components in advanced online VIS architecture (i.e., transformer-based architecture), we find that the transformer decoder significantly hampers the inference speed. Further analysis of the similarities between the outputs from adjacent frames at each transformer decoder layer reveals significant redundant computations within the transformer decoder. To address this issue, we introduce Temporal-Aware query Routing (TAR) mechanism. We embed it before each transformer decoder layer. By fusing the optimal queries from the previous frame, the queries output by the preceding decoder layer, and their differential information, TAR predicts a binary classification score and then uses an argmax operation to determine whether the current layer should be skipped. Experimental results demonstrate that integrating TAR into the baselines achieves significant efficiency gains (24.7 - 34.6 FPS for MinVIS, 22.4 - 32.8 FPS for DVIS++) while also improving performance (e.g., on YoutubeVIS 2019, 47.4 - 48.4 AP for MinVIS, 55.5 - 55.7 AP for DVIS++). Furthermore, our analysis of the TAR mechanism shows that the number of skipped layers increases as the differences between adjacent video frames decrease, which suggests that our method effectively utilizes inter-frame differences to reduce redundant computations in the transformer decoder.

With the rise of social media, vast amounts of user-uploaded videos (e.g., YouTube) are utilized as training data for Visual Object Tracking (VOT). However, the VOT community has largely overlooked video data-privacy issues, as many private videos have been collected and used for training commercial models without authorization. To alleviate these issues, this paper presents the first investigation on preventing personal video data from unauthorized exploitation by deep trackers. Existing methods for preventing unauthorized data use primarily focus on image-based tasks (e.g., image classification), directly applying them to videos reveals several limitations, including inefficiency, limited effectiveness, and poor generalizability. To address these issues, we propose a novel generative framework for generating Temporal Unlearnable Examples (TUEs), and whose efficient computation makes it scalable for usage on large-scale video datasets. The trackers trained w/ TUEs heavily rely on unlearnable noises for temporal matching, ignoring the original data structure and thus ensuring training video data-privacy. To enhance the effectiveness of TUEs, we introduce a temporal contrastive loss, which further corrupts the learning of existing trackers when using our TUEs for training. Extensive experiments demonstrate that our approach achieves state-of-the-art performance in video data-privacy protection, with strong transferability across VOT models, datasets, and temporal matching tasks.

Human Motion Segmentation (HMS), which aims to partition videos into non-overlapping human motions, has attracted increasing research attention recently. Existing approaches for HMS are mainly dominated by subspace clustering methods, which are grounded on the assumption that high-dimensional temporal data align with a Union-of-Subspaces (UoS) distribution. However, the frames in video capturing complex human motions with cluttered backgrounds may not align well with the UoS distribution. In this paper, we propose a novel approach for HMS, named Temporal Rate Reduction Clustering (\text TR ^2\text C ), which jointly learns structured representations and affinity to segment the sequences of frames in video. Specifically, the structured representations learned by \text TR ^2\text C enjoy temporally consistency and are aligned well with a UoS structure, which is favorable for addressing the HMS task. We conduct extensive experiments on five benchmark HMS datasets and achieve state-of-the-art performances with different feature extractors. The code is available at: https://github.com/mengxianghan123/TR2C.

Moving object segmentation (MOS) on LiDAR point clouds is crucial for autonomous systems such as self-driving vehicles. While previous supervised approaches rely on costly manual annotations, LiDAR sequences naturally capture temporal motion cues that can be leveraged for self-supervised learning. In this paper, we propose Temporal Overlapping Prediction (TOP), a self-supervised pre-training method designed to alleviate this annotation burden. TOP learns powerful spatiotemporal representations by predicting the occupancy states of temporal overlapping points that are commonly observed in current and adjacent scans. To further ground these representations in the current scene's geometry, we introduce an auxiliary pre-training objective of reconstructing the occupancy of the current scan. Extensive experiments on the nuScenes and SemanticKITTI datasets validate our method's effectiveness. TOP consistently outperforms existing supervised and self-supervised pre-training baselines across both point-level Intersection-over-Union (IoU) and object-level Recall metrics. Notably, it achieves a relative improvement of up to 28.77% over a training-from-scratch baseline and demonstrates strong transferability across LiDAR setups. Our code is publicly available at https://github.com/ZiliangMiao/TOP.

While deep models are effectively trained based on a softmax cross-entropy loss, a cosine-based softmax loss also works for producing favorable feature embedding.In the cosine-based softmax, temperature plays a crucial role in properly scaling the logits of cosine similarities, though being manually tuned in ad-hoc ways as there is less prior knowledge about the temperature.In this paper, we address the challenging problem to adaptively estimate the temperature of cosine-based softmax in the framework of supervised image classification.By analyzing the cosine-based softmax representation from a geometrical viewpoint regarding features and classifiers, we construct a criterion in a least-square fashion which enables us to optimize the temperature at each sample via simple greedy search.Besides, our thorough analysis about temperature clarifies that feature embedding by the cosine-based softmax loss is endowed with diverse characteristics which are controllable by the temperature in an explainable way.The experimental results demonstrate that our optimized temperature contributes to determine a feasible range of temperature to control the feature characteristics and produces favorable performance on various image classification tasks.

The task of realistically inserting a human from a reference image into a background scene is highly challenging, requiring the model to (1) determine the correct location and poses of the person and (2) perform high-quality personalization conditioned on the background. Previous approaches often treat them as separate problems, overlooking their interconnections, and typically rely on training to achieve high performance. In this work, we introduce a unified training-free pipeline that leverages pre-trained text-to-image diffusion models. We show that diffusion models inherently possess the knowledge to place people in complex scenes without requiring task-specific training. By combining inversion techniques with classifier-free guidance, our method achieves affordance-aware global editing, seamlessly inserting people into scenes. Furthermore, our proposed mask-guided self-attention mechanism ensures high-quality personalization, preserving the subject's identity, clothing, and body features from just a single reference image. To the best of our knowledge, we are the first to perform realistic human insertions into scenes in a training-free manner and achieve state-of-the-art results in diverse composite scene images with excellent identity preservation in backgrounds and subjects.

We present the first method for personalized dental shape reconstruction and teeth-inclusive facial performance capture using only a single phone camera. Our approach democratizes high-quality facial avatars through a non-invasive, low-cost setup by addressing the ill-posed monocular capture problem with an analysis-by-synthesis approach. We introduce a representation adaptation technique that maintains both mesh and SDF representations of teeth, enabling efficient differentiable rendering while preventing teeth-lip interpenetration. To overcome alignment challenges with similar-appearing dental components, we leverage foundation models for semantic teeth segmentation and design specialized optimization objectives. Our method addresses the challenging occlusions of teeth during facial performance through optimization strategies that leverage facial structural priors, while our semantic mask rendering loss with optimal transport-based matching ensures convergence despite significant variations in initial positioning.

Vision-Language Models (VLMs) have shown remarkable capabilities across diverse visual tasks, including image recognition, video understanding, and Visual Question Answering (VQA) when explicitly trained for these tasks. Despite these advances, we find that present-day VLMs (including the proprietary GPT-4o) lack a fundamental cognitive ability: learning to localize specific objects in a scene by taking into account the context.In this work, we focus on the task of few-shot personalized localization, where a model is given a small set of annotated images (in-context examples) -- each with a category label and bounding box -- and is tasked with localizing the same object type in a query image. Personalized localization can be particularly important in cases of ambiguity of several related objects that can respond to a text or an object that is hard to describe with words.To provoke personalized localization abilities in models, we present a data-centric solution that fine-tunes them using carefully curated data from video object tracking datasets. By leveraging sequences of frames tracking the same object across multiple shots, we simulate instruction-tuning dialogues that promote context awareness. To reinforce this, we introduce a novel regularization technique that replaces object labels with pseudo-names, ensuring the model relies on visual context rather than prior knowledge. Our method significantly enhances few-shot localization performance of recent VLMs ranging from 7B to 72B in size, without sacrificing generalization, as demonstrated on several benchmarks tailored towards evaluating personalized localization abilities. This work is the first to explore and benchmark personalized few-shot localization for VLMs -- exposing critical weaknesses in present-day VLMs, and lays a foundation for future research in context-driven vision-language applications.

Imposing key anatomical features, such as the number of organs, their shapes and relative positions, is crucial for building a robust multi-organ segmentation model. Current attempts to incorporate anatomical features include broadening the effective receptive field (ERF) size with data-intensive modules, or introducing anatomical constraints that scales poorly to multi-organ segmentation. We introduce a novel architecture called the Anatomy-Informed Cascaded Segmentation Network (AIC-Net). AIC-Net incorporates a learnable input termed "Anatomical Prior", which can be adapted to patient-specific anatomy using a differentiable spatial deformation. The deformed prior later guides decoder layers towards more anatomy-informed predictions. We repeat this process at a local patch level to enhance the representation of intricate objects, resulting in a cascaded network structure. AIC-Net is a general method that enhances any existing segmentation models to be more anatomy-aware. We have validated the performance of AIC-Net, with various backbones, on three multi-organ segmentation tasks: abdominal organs, vertebrae, and ribs. For each respective task, our benchmarks demonstrate improved dice score and Hausdorff distance.

Efficient 3D avatar creation is a significant demand in the metaverse, film/game, AR/VR, etc. In this paper, we rethink text-to-avatar generative models by proposing TeRA, a more efficient and effective framework than the previous SDS-based models and general large 3D generative models. Our approach employs a two-stage training strategy for learning a native 3D avatar generative model. Initially, we distill a decoder to derive a structured latent space from a large human reconstruction model. Subsequently, a text-controlled latent diffusion model is trained to generate photorealistic 3D human avatars within this latent space. TeRA enhances the model performance by eliminating slow iterative optimization and enables text-based partial customization through a structured 3D human representation. Experiments have proven our approach's superiority over previous text-to-avatar generative models in subjective and objective evaluation.

Recent advances in text-to-image synthesis largely benefit from sophisticated sampling strategies and classifier-free guidance (CFG) to ensure high-quality generation. However, CFG's reliance on two forward passes, especially when combined with intricate sampling algorithms, results in prohibitively high inference costs. To address this, we introduce TeEFusion (Text Embeddings Fusion), a novel and efficient distillation method that directly incorporates the guidance magnitude into the text embeddings and distills the teacher model's complex sampling strategy. By simply fusing conditional and unconditional text embeddings using linear operations, TeEFusion reconstructs the desired guidance without adding extra parameters, simultaneously enabling the student model to learn from the teacher's output produced via its sophisticated sampling approach. Extensive experiments on state-of-the-art models such as SD3 demonstrate that our method allows the student to closely mimic the teacher's performance with a far simpler and more efficient sampling strategy. Consequently, the student model achieves inference speeds up to 6xfaster than the teacher model, while maintaining image quality at levels comparable to those obtained through the teacher's complex sampling approach. The code is publicly available at github.com/AIDC-AI/TeEFusion.

We propose TaxaDiffusion, a taxonomy-informed training framework for diffusion models to generate fine-grained animal images with high morphological and identity accuracy. Unlike standard approaches that treat each species as an independent category, TaxaDiffusion incorporates domain knowledge that many species exhibit strong visual similarities, with distinctions often residing in subtle variations of shape, pattern, and color. To exploit these relationships, TaxaDiffusion progressively trains conditioned diffusion models across different taxonomic levels --- starting from broad classifications such as Class and Order, refining through Family and Genus, and ultimately distinguishing at the Species level. This hierarchical learning strategy first captures coarse-grained morphological traits shared by species with common ancestors, facilitating knowledge transfer, before refining fine-grained differences for species-level distinction. As a result, TaxaDiffusion enables accurate generation even with limited training samples per species. Extensive experiments on three fine-grained animal datasets demonstrate that TaxaDiffusion outperforms existing approaches, achieving superior fidelity in fine-grained animal image generation. Our model and code will be publicly available.