ICCV 2025 Papers

Monocular 3D object detectors, while effective on data from one ego camera height, struggle with unseen or out-of-distribution camera heights. Existing methods often rely on Plucker embeddings, image transformations or data augmentation. This paper takes a step towards this understudied problem by investigating the impact of camera height variations on state-of-the-art (SoTA) Mono3D models. With a systematic analysis on the extended CARLA dataset with multiple camera heights, we observe that depth estimation is a primary factor influencing performance under height variations. We mathematically prove and also empirically observe consistent negative and positive trends in mean depth error of regressed and ground-based depth models, respectively, under camera height changes. To mitigate this, we propose Camera Height Robust Monocular 3D Detector (CHARM3R), which averages both depth estimates within the model. CHARM3R significantly improves generalization to unseen camera heights, by more than 45%, achieving SoTA performance on the CARLA dataset.

3D Gaussian Splatting (3DGS) has begun incorporating rich information from 2D foundation models. However, most approaches rely on a bottom-up optimization process that treats raw 2D features as ground truth, incurring increased computational costs. We propose a top-down pipeline for constructing compact and fast 3D feature fields, namely, CF3. We first perform a fast weighted fusion of multi-view 2D features with pre-trained Gaussians. This approach enables training a per-Gaussian autoencoder directly on the lifted features, instead of training autoencoders in the 2D domain. As a result, the autoencoder better aligns with the inference-time feature distribution, achieving high compression fidelity and storage efficiency. Finally, we introduce an adaptive sparsification method that prunes and merges Gaussians, constructing an efficient representation while maintaining details. Our approach achieves a competitive 3D feature field while using as little as 5% of the Gaussians compared to Feature-3DGS.

Although saliency maps can highlight important regions to explain the reasoning behind image classification in artificial intelligence (AI), the meaning of these regions is left to the user's interpretation. In contrast, concept-based explanations decompose AI predictions into human-understandable concepts, clarifying their contributions. However, few methods can simultaneously reveal what concepts an image classifier learns, which regions are associated with them, and how they contribute to predictions.We propose a novel concept-based explanation method, Concept-based Explanation via Fusion of Activation Maps (CE-FAM). It employs a branched network that shares activation maps with an image classifier and learns to mimic the embeddings of a Vision and Language Model (VLM). The branch network predicts concepts in an image, and their corresponding regions are represented by a weighted sum of activation maps, with weights given by the gradients of the concept prediction scores. Their contributions are quantified based on their impact on the image classification score. Our method provides a general framework for identifying the concept regions and their contributions while leveraging VLM knowledge to handle arbitrary concepts without requiring an annotated dataset. Furthermore, we introduce a novel evaluation metric to assess the accuracy of the concept regions. Our qualitative and quantitative evaluations demonstrate our method outperforms existing approaches and excels in zero-shot inference for unseen concepts.

Computational color constancy, or white balancing, is a key module in a camera's image signal processor (ISP) that corrects color casts from scene lighting. Because this operation occurs in the camera-specific raw color space, white balance algorithms must adapt to different cameras. This paper introduces a learning-based method for cross-camera color constancy that generalizes to new cameras without retraining. Our method leverages pre-calibrated color correction matrices (CCMs) available on ISPs that map the camera's raw color space to a standard space (e.g., CIE XYZ). Our method uses these CCMs to transform predefined illumination colors (i.e., along the Planckian locus) into the test camera's raw space. The mapped illuminants are encoded into a compact camera fingerprint embedding (CFE) that enables the network to adapt to unseen cameras. To prevent overfitting due to limited cameras and CCMs during training, we introduce a data augmentation technique that interpolates between cameras and their CCMs. Experimental results across multiple datasets and backbones show that our method achieves state-of-the-art cross-camera color constancy while remaining lightweight and relying only on data readily available in camera ISPs.

Recent advances in 3D reconstruction techniques and vision-language models have fueled significant progress in 3D semantic understanding, a capability critical to robotics, autonomous driving, and virtual/augmented reality. However, methods that rely on 2D priors are prone to a critical challenge: cross-view semantic inconsistencies induced by occlusion, image blur, and view-dependent variations. These inconsistencies, when propagated via projection supervision, deteriorate the quality of 3D Gaussian semantic fields and introduce artifacts in the rendered outputs. To mitigate this limitation, we propose CCL-LGS, a novel framework that enforces view-consistent semantic supervision by integrating multi-view semantic cues. Specifically, our approach first employs a zero-shot tracker to align a set of SAM-generated 2D masks and reliably identify their corresponding categories. Next, we utilize CLIP to extract robust semantic encodings across views. Finally, our Contrastive Codebook Learning (CCL) module distills discriminative semantic features by enforcing intra-class compactness and inter-class distinctiveness. In contrast to previous methods that directly apply CLIP to imperfect masks, our framework explicitly resolves semantic conflicts while preserving category discriminability. Extensive experiments demonstrate that CCL-LGS outperforms previous state-of-the-art methods. Our project page is available at https://epsilontl.github.io/CCL-LGS/.

Large Multimodal Models (LMMs) have demonstrated impressive performance in recognizing document images with natural language instructions. However, it remains unclear to what extent capabilities in literacy with rich structure and fine-grained visual challenges. The current landscape lacks a comprehensive benchmark to effectively measure the literate capabilities of LMMs. Existing benchmarks are often limited by narrow scenarios and specified tasks. To this end, we introduce CC-OCR, a comprehensive benchmark that possesses a diverse range of scenarios, tasks, and challenges. CC-OCR comprises four OCR-centric tracks: multi-scene text reading, multilingual text reading, document parsing, and key information extraction. It includes 39 subsets with 7,058 full annotated images, of which 41% are sourced from real applications, and released for the first time. We evaluate ten prominent LMMs and reveal both the strengths and weaknesses of these models, particularly in text grounding, multi-orientation, and hallucination of repetition. CC-OCR aims to comprehensively evaluate the capabilities of LMMs on OCR-centered tasks, facilitating continued progress in this crucial area.

In this paper, we introduce the Context-Aware Video Instance Segmentation (CAVIS), a novel framework designed to enhance instance association by integrating contextual information adjacent to each object. To efficiently extract and leverage this information, we propose the Context-Aware Instance Tracker (CAIT), which merges contextual data surrounding the instances with the core instance features to improve tracking accuracy. Additionally, we design the Prototypical Cross-frame Contrastive (PCC) loss, which ensures consistency in object-level features across frames, thereby significantly enhancing instance matching accuracy. CAVIS demonstrates superior performance over state-of-the-art methods on all benchmark datasets in video instance segmentation (VIS) and video panoptic segmentation (VPS). Notably, our method excels on the OVIS dataset, which is known for its particularly challenging videos.

Recently, generalizable feed-forward methods based on 3D Gaussian Splatting have gained significant attention for their potential to reconstruct 3D scenes using finite resources. These approaches create a 3D radiance field, parameterized by per-pixel 3D Gaussian primitives, from just a few images in a single forward pass. Unlike multi-view methods that benefit from cross-view correspondences, 3D scene reconstruction with a single-view image remains an underexplored area. In this work, we introduce CATSplat, a novel generalizable transformer-based framework designed to break through the inherent constraints in monocular settings. First, we propose leveraging textual guidance from a visual-language model to complement insufficient information from single-view image features. By incorporating scene-specific contextual details from text embeddings through cross-attention, we pave the way for context-aware 3D scene reconstruction beyond relying solely on visual cues. Moreover, we advocate utilizing spatial guidance from 3D point features toward comprehensive geometric understanding under monocular settings. With 3D priors, image features can capture rich structural insights for predicting 3D Gaussians without multi-view techniques. Extensive experiments on large-scale datasets demonstrate the state-of-the-art performance of CATSplat in single-view 3D scene reconstruction with high-quality novel view synthesis.

Utilizing large language models (LLMs) for tool planning has emerged as a promising avenue for developing general AI systems, where LLMs automatically schedule external tools (e.g., vision models) to tackle complex tasks based on task descriptions. To push this paradigm toward practical applications, it is crucial for LLMs to consider tool execution costs (e.g., execution time) for tool planning. Unfortunately, prior studies overlook the tool execution costs, leading to the generation of expensive plans whose costs outweigh their benefits in terms of task performance. To fill this gap, we propose the Cost-Aware Tool Planning with LLMs (CATP-LLM) framework, which for the first time provides a coherent design to empower LLMs for cost-aware tool planning. Specifically, To facilitate efficient concurrent tool execution and cost reduction, we design a tool planning language to enhance the LLM for creating multi-branch non-sequential plans. Moreover, we propose a cost-aware offline reinforcement learning algorithm to fine-tune the LLM to optimize the performance-cost trade-off in tool planning. In the lack of public cost-related datasets, we further present OpenCATP, the first dataset for cost-aware planning, which comprises 11,100 evaluation samples from diverse tasks. Extensive experiments show that CATP-LLM outperforms GPT-4 even when using Llama2-7B as its backbone, with the average plan quality improvement of 1.5%-93.9%. Codes and dataset are available at: https://github.com/duowuyms/OpenCATP-LLM.

Modern visual trackers have achieved robust performance with precisely initialized target bounding boxes. However, providing high-precision initial annotations is a process both labor-intensive and error-prone in real-world scenarios. Interactive initialization (e.g., click-based, scribble-based) presents a more practical alternative. In this paper, we introduce a unified Click-and-Track (CAT) framework for full-process tracking, eliminating the need for auxiliary models or complex initialization pipelines. We present a novel fine-tuning paradigm that bridges the information gap inherent in click-based initialization through two key innovations: 1) The proposed click-based localization and joint spatial-visual prompt refinement are sequentially performed to compensate for the geometric information loss (e.g., boundary ambiguity, shape uncertainty) inherent in click-based initialization. 2) We design a parameter-efficient module called CTMoE to leverage the tracker's inherent capabilities when fine-tuning. The proposed CTMoE enables the foundation model to learn different matching patterns, unifying click-based initialization and tracking within a unified architecture. Extensive experimental results demonstrate state-of-the-art performance of our click-based tracking method on the LaSOT benchmark (70.5% AUC) while maintaining parameter efficiency, surpassing existing click-based tracking frameworks by a large margin and even outperforming some bounding-box-initialized trackers. The code and models are available at https://github.com/ysyuann/CAT.

In robotic visuomotor policy learning, diffusion-based models have achieved significant success in improving the accuracy of action trajectory generation compared to traditional autoregressive models. However, they suffer from inefficiency due to multiple denoising steps and limited flexibility from complex constraints. In this paper, we introduce **C**oarse-to-Fine **A**uto**R**egressive **P**olicy (**CARP**), a novel paradigm for visuomotor policy learning that redefines the autoregressive action generation process as a coarse-to-fine, next-scale approach. CARP decouples action generation into two stages: first, an action autoencoder learns multi-scale representations of the entire action sequence; then, a GPT-style transformer refines the sequence prediction through a coarse-to-fine autoregressive process. This straightforward and intuitive approach produces highly accurate and smooth actions, matching or even surpassing the performance of diffusion-based policies while maintaining efficiency on par with autoregressive policies. We conduct extensive evaluations across diverse settings, including single-task and multi-task scenarios on state-based and image-based simulation benchmarks, as well as real-world tasks. CARP achieves competitive success rates, with up to a 10% improvement, and delivers **10x** faster inference compared to state-of-the-art policies, establishing a high-performance, efficient, and flexible paradigm for action generation in robotic tasks.

Performance predictors have emerged as a promising method to accelerate the evaluation stage of neural architecture search (NAS). These predictors estimate the performance of unseen architectures by learning from the correlation between a small set of trained architectures and their performance. However, most existing predictors ignore the inherent distribution shift between limited training samples and diverse test samples. Hence, they tend to learn spurious correlations as shortcuts to predictions, leading to poor generalization. To address this, we propose a Causality-guided Architecture Representation Learning (CARL) method aiming to separate critical (causal) and redundant (non-causal) features of architectures for generalizable architecture performance prediction. Specifically, we employ a substructure extractor to split the input architecture into critical and redundant substructures in the latent space. Then, we generate multiple interventional samples by pairing critical representations with diverse redundant representations to prioritize critical features. Extensive experiments on five NAS search spaces demonstrate the state-of-the-art accuracy and superior interpretability of CARL. For instance, CARL achieves 97.67% top-1 accuracy on CIFAR-10 using DARTS.

Text-to-video retrieval serves as a powerful tool for navigating vast video databases. This is particularly useful in autonomous driving to retrieve scenes from a text query to simulate and evaluate the driving system in desired scenarios. However, traditional ranking-based retrieval methods often return partial matches that do not satisfy all query conditions. To address this, we introduce Inclusive Text-to-Video Retrieval, which retrieves only videos that meet all specified conditions, regardless of additional irrelevant elements. We propose CARIM, a framework for driving scene retrieval that employs inclusive text matching. By utilizing Vision-Language Model (VLM) and Large Language Model (LLM) to generate compressed captions for driving scenes, we transform text-to-video retrieval into a more efficient text-to-text retrieval problem, eliminating modality mismatches and heavy annotation costs. We introduce a novel positive and negative data curation strategy and an attention-based scoring mechanism tailored for driving scene retrieval. Experimental results on the DRAMA dataset demonstrate that CARIM outperforms state-of-the-art retrieval methods, excelling in edge cases where traditional models fail.

Recognizing and reasoning about occluded (partially or fully hidden) objects is vital to understanding visual scenes, as occlusions frequently occur in real-world environments and act as obstacles for spatial comprehension. To test models' ability to reason about multiple occluded objects, we introduce a novel task, Counting Amodally for Patterns Through Unseen REgions (CAPTURe), which requires a model to count objects arranged in a pattern by inferring how the pattern continues behind an occluder (an object which blocks parts of the scene). CAPTURe requires both recognizing visual patterns and reasoning, making it a useful testbed for evaluating vision-language models (VLMs) on whether they understand occluded patterns and possess spatial understanding skills. By requiring models to reason about occluded objects, CAPTURe also tests VLMs' ability to form world models that would allow them to fill in missing information. CAPTURe consists of two parts: (1) CAPTURe-real, with manually filtered images of real objects in patterns and (2) CAPTURe-synthetic, a controlled diagnostic with generated patterned images. We evaluate four strong VLMs (GPT-4o, Intern-VL2, Molmo, and Qwen2-VL) on CAPTURe, finding that models struggle to count on both occluded and unoccluded patterns. Crucially, we find that models perform worse with occlusion, suggesting that VLMs are also deficient in inferring unseen spatial relationships: even the strongest VLMs like GPT-4o fail to count with occlusion. In contrast, we find that humans achieve very little error on CAPTURe. We also find that providing auxiliary information of occluded object locations increases performance, underscoring that the model error comes both from an inability to handle occlusion as well as difficulty in counting in images.

We address the task of advertisement image generation and introduce three evaluation metrics to assess Creativity, prompt Alignment, and Persuasiveness (CAP) in generated advertisement images. Despite recent advancements in Text-to-Image (T2I) methods and their performance in generating high-quality images for explicit descriptions, evaluating these models remains challenging. Existing evaluation methods focus largely on assessing alignment with explicit, detailed descriptions, but evaluating alignment with visually implicit prompts remains an open problem. Additionally, creativity and persuasiveness are essential qualities that enhance the effectiveness of advertisement images, yet are seldom measured. To address this, we propose three novel metrics for evaluating the creativity, alignment, and persuasiveness of generated images. We show that current T2I models struggle with creativity, persuasiveness, and alignment when the input text is implicit messages. We further introduce a simple yet effective approach to enhance T2I models' capabilities in producing images that are better aligned, more creative, and more persuasive.

Foley is a key element in video production, refers to the process of adding an audio signal to a silent video while ensuring semantic and temporal alignment. In recent years, the rise of personalized content creation and advancements in automatic video-to-audio models have increased the demand for greater user control in the process. One possible approach is to incorporate text to guide audio generation. While supported by existing methods, challenges remain in ensuring compatibility between modalities, particularly when the text introduces additional information or contradicts the sounds naturally inferred from the visuals. In this work, we introduce CAFA (Controllable Automatic Foley Artist) a video-and-text-to-audio model that generates semantically and temporally aligned audio for a given video, guided by text input. CAFA is built upon a text-to-audio model and integrates video information through a modality adapter mechanism. By incorporating text, users can refine semantic details and introduce creative variations, guiding the audio synthesis beyond the expected video contextual cues. Experiments show that besides its superior quality in terms of semantic alignment and audio-visual synchronization the proposed method enable high textual controllability as demonstrated in subjective and objective evaluations.

Computer-Aided Design (CAD) models are typically constructed by sequentially drawing parametric sketches and applying CAD operations to obtain a 3D model. The problem of 3D CAD reverse engineering consists of reconstructing the sketch and CAD operation sequences from 3D representations such as point clouds. In this paper, we address this challenge through novel contributions across three levels: CAD sequence representation, network design, and training dataset. In particular, we represent CAD sketch-extrude sequences as Python code. The proposed CAD-Recode translates a point cloud into Python code that, when executed, reconstructs the CAD model. Taking advantage of the exposure of pre-trained Large Language Models (LLMs) to Python code, we leverage a relatively small LLM as a decoder for CAD-Recode and combine it with a lightweight point cloud projector. CAD-Recode is trained on a procedurally generated dataset of one million CAD sequences. CAD-Recode significantly outperforms existing methods across the DeepCAD, Fusion360 and real-world CC3D datasets. Furthermore, we show that our CAD Python code output is interpretable by off-the-shelf LLMs, enabling CAD editing and CAD-specific question answering from point clouds.

We propose CAD-Assistant, a general-purpose CAD agent for AI-assisted design. Our approach is based on a powerful Vision and Large Language Model (VLLM) as a planner and a tool-augmentation paradigm using CAD-specific tools. CAD-Assistant addresses multimodal user queries by generating actions that are iteratively executed on a Python interpreter equipped with the FreeCAD software, accessed via its Python API. Our framework is able to assess the impact of generated CAD commands on geometry and adapts subsequent actions based on the evolving state of the CAD design. We consider a wide range of CAD-specific tools including a sketch image parameterizer, rendering modules, a 2D cross-section generator, and other specialized routines. CAD-Assistant is evaluated on multiple CAD benchmarks, where it outperforms VLLM baselines and supervised task-specific methods. Beyond existing benchmarks, we qualitatively demonstrate the potential of tool-augmented VLLMs as general-purpose CAD solvers across diverse workflows.