ICCV 2025 · Paper Collection
ICCV 2025 object pose research lowers unseen-object onboarding cost while incorporating environmental decoupling, structured 3D representations, result reliability, and industrial precision into the same technical chain.
12
Papers in this collection
12
Official Posters
4
Unseen objects / few references
3
Robustness and deployment studies
These 12 papers reflect this site's ICCV 2025 main-conference inclusion criteria, not an official task-category count. Titles, authors, and presentation types follow the official accepted-paper directory and Open Access pages. All included papers are Posters; no official Highlight / Oral labels were found.
Collection in One Sentence
6DOPE-GS, MixRI, RayPose, and BoxDreamer address object onboarding: rapidly converting known CAD, sparse references, or online RGB-D observations into usable query poses. Their intermediate representations include explicit correspondences, ray diffusion, box corners, and Gaussian object fields.
SDFit, CleanPose, and Unified Detection & Pose incorporate valid shape manifolds, causal bias, and 3D prototypes into inference. EA6D, deterministic confidence regions, and ultraprecise measurement explicitly address when predictions can be trusted and under which conditions they work.
Research Trends
These themes summarize the 12 included main-conference papers. Papers may have multiple tags, but primary groups follow their main problem settings.
01 · Object Onboarding
MixRI fuses point features across reference views, RayPose recasts templates as ray alignment, and BoxDreamer compresses sparse views into stable bounding-box corners. All aim to reduce new-object reference acquisition and initialization cost.
02 · Structured 3D Reasoning
SDFit jointly fits shape and pose within a deformable SDF manifold; Unified Detection & Pose unifies detection and pose through neural-mesh prototypes; CleanPose explains cross-instance failures through data bias and causal variables.
03 · Reliability by Design
EA6D separates environmental factors such as lighting, backgrounds, and occlusion; 3D compositional models investigate OOD conditions; Deterministic Pose Confidence propagates keypoint regions into 6D pose confidence regions at user-specified coverage.
04 · Beyond Benchmark Settings
CapeLLM replaces traditional support keypoints with query images and text descriptions, while Ultra-Precision 6DoF targets periodic patterns, nanometer translation, and microradian rotation in industrial measurement. Their scores are not directly comparable with BOP, but they reveal boundaries of pose systems.
Data Statistics
Each paper belongs to one primary-problem group; overlapping methodological attributes remain in its tags.
Builds correspondences, templates, and online object representations for objects unseen during training using CAD or a few posed reference images.
Improves joint pose, shape, and detection inference across instances through deformable shapes, neural meshes, and causal / 3D priors.
Treats environmental perturbations, out-of-distribution generalization, result reliability, and ultraprecise measurement as primary pose-system objectives.
Explores text, vision–language models, and spatial reasoning to reduce support-image requirements and explicit category priors.
Interpreting the statistics: 4 papers directly address unseen objects or few reference views, 4 use structured shapes / representations for category-level inference, and 3 treat environment, confidence, or measurement precision as core variables. CapeLLM explores category-agnostic keypoint pose and is listed separately from standard 6D pose benchmarks.
Recommended Reading
Recommendations reflect problem coverage, methodological approaches, and engineering insights, rather than official ICCV selections or rankings.
Connects model-free object tracking, Gaussian Splatting representations, and online reconstruction in a closed loop for real-time applications.
Targets practical constraints of fewer references and lighter models, providing a useful engineering comparison for BOP approaches.
Reformulates template matching as camera-ray alignment and applies a diffusion transformer, representing a clear direction in generative pose representations.
Combines an interpretable render-and-compare feedback loop with deformable SDFs and foundation-feature initialization, offering insights into single-image pose-and-shape system design.
Unifies detection and pose through one 3D prototype representation, addressing fragmented RGB category-level deployment pipelines.
Identifies data bias as a category-generalization bottleneck, using front-door adjustment and 3D foundation-model distillation to address causal and semantic biases, respectively.
Extends pose predictions to calibrated confidence regions, closely matching the needs of grasping, aerospace, and high-reliability systems.
Although distinct from BOP-style 6D pose, it brings MLLM language and spatial priors into category-agnostic pose, making it relevant as a boundary exploration of reduced explicit references.
About Highlight / Oral designations: At this review's check against the official ICCV 2025 accepted-paper directory and paper pages, all 12 entries are labeled Poster. Relative camera pose, human pose, scene reconstruction alone, and workshop papers are excluded; site recommendation labels do not replace official presentation designations.
Paper List
Each card provides task boundaries, input modalities, tags, and a short summary. Select a card for the site's coarse and deep notes; the official ICCV source remains linked at the lower right.
4 papers
Builds correspondences, templates, and online object representations for objects unseen during training using CAD or a few posed reference images.
Yufeng Jin · Vignesh Prasad · Snehal Jauhri · Mathias Franzius · Georgia Chalvatzaki
Setting: Online 6D pose tracking and reconstruction of unknown objects
Input: Monocular RGB-D video
Summary: Jointly optimizes object poses and representations using incremental 2D Gaussian Splatting. Dynamic keyframe selection and opacity-statistics pruning reduce incorrect updates and training cost, emphasizing real-time performance on HO3D and YCBInEOAT.
Xinhang Liu · Jiawei Shi · Zheng Dang · Yuchao Dai
Setting: Novel-object 6D pose with CAD / rendered references
Input: RGB
Summary: Fuses features of the same 3D point across references through View-Aggregated Point Matching, directly predicts query correspondences and occlusion, then solves pose through RANSAC-PnP. Aims for competitive performance on 7 BOP core datasets with fewer references and less computation.
Junwen Huang · Shishir Reddy Vutukur · Peter KT Yu · Nassir Navab · Slobodan Ilic · Benjamin Busam
Setting: 6D pose from multiple posed template views to an unseen query
Input: RGB
Summary: Represents rotation with object-centric camera rays and translation with scale-invariant dense offsets. Conditioned on posed template sets, the model recovers query pose through coarse-to-fine template sampling.
Yuanhong Yu · Xingyi He · Chen Zhao · Junhao Yu · Jiaqi Yang · Ruizhen Hu · Yujun Shen · Xing Zhu · Xiaowei Zhou · Sida Peng
Setting: Generalizable object pose from sparse RGB references
Input: RGB + sparse reference views
Summary: First recovers an object 3D bounding box from sparse reference views, then predicts query-image 2D box corners with a reference-based point synthesizer. These corners naturally provide 2D–3D correspondences for PnP, improving robustness to occlusion and sparse viewpoints.
4 papers
Improves joint pose, shape, and detection inference across instances through deformable shapes, neural meshes, and causal / 3D priors.
Dimitrije Antić · Georgios Paschalidis · Shashank Tripathi · Theo Gevers · Sai Kumar Dwivedi · Dimitrios Tzionas
Setting: Category-level 3D pose and shape from a single natural image
Input: RGB
Summary: Constrains valid shapes through a category-specific deformable SDF manifold. Retrieves a similar 3D shape and establishes rich 2D–3D correspondences for pose initialization, then jointly optimizes pose and shape. Emphasizes generalization without retraining under occlusion, unusual poses, and in-the-wild images.
Tom Fischer · Xiaojie Zhang · Eddy Ilg
Setting: Joint category-level multi-object detection and pose from RGB
Input: RGB
Summary: Uses neural meshes with learned features as category 3D prototypes and multi-model RANSAC to detect objects and estimate pose in one framework. Reports substantial gains on REAL275 RGB category-level scale-agnostic metrics.
Xiao Lin · Yun Peng · Liuyi Wang · Xianyou Zhong · Minghao Zhu · Yi Feng · Jingwei Yang · Chengju Liu · Qijun Chen
Setting: Category-level object pose and size with reduced data bias
Input: RGB-D
Summary: Builds a causal inference module from structural information and hidden confounders, reducing spurious correlations through front-door adjustment. Residual distillation transfers unbiased semantics from a 3D foundation model; evaluated on REAL275, CAMERA25, and HouseCat6D.
Xiaoding Yuan · Guofeng Zhang · Prakhar Kaushik · Artur Jesslen · Adam Kortylewski · Alan Yuille
Setting: Joint classification and 3D pose inference across many categories
Input: RGB
Summary: Decomposes cross-category neural-mesh vertex contrastive learning into within-class and between-class components and dynamically focuses on confusable categories, scaling structured models to 188 categories. Evaluates OOD robustness to occlusion, weather, synthetic data, and unknown categories.
3 papers
Treats environmental perturbations, out-of-distribution generalization, result reliability, and ultraprecise measurement as primary pose-system objectives.
Shaobo Zhang · Yuhang Huang · Wanqing Zhao · Wei Zhao · Ziyu Guan · Jinye Peng
Setting: 6D object pose under environmental changes
Input: RGB
Summary: EA6D uses an Environment Decoupling Diffusion Model to generate object representations decoupled from lighting, background, and occlusion using input-image priors, with few-step sampling for efficiency. Trains only on synthetic data and evaluates on 4 standard benchmarks and DiverseScenes.
Jinghao Wang · Zhang Li · Zi Wang · Banglei Guan · Yang Shang · Qifeng Yu
Setting: 6D pose confidence regions and safety quantification
Input: RGB keypoint pipeline
Summary: Calibrates Gaussian keypoint distributions from deterministic regression into 2D regions through inductive conformal prediction, then propagates them to 6D pose regions using the implicit function theorem, avoiding slow sampling / ensembles and inflated regions.
Guowei Shi · Zian Mao · Peisen Huang
Setting: Ultraprecise 6DoF measurement from periodic patterns
Input: Industrial camera + etched checkerboard pattern
Summary: Introduces a 2D interpolated DFT and a mathematical mapping from image phase/frequency to 6DoF pose, supporting orthographic and near-orthographic imaging. Low-cost industrial-camera experiments report nanometer-level translation and microradian-level rotation precision.
1 papers
Explores text, vision–language models, and spatial reasoning to reduce support-image requirements and explicit category priors.
Junho Kim · Hyungjin Chung · Byung-Hoon Kim
Setting: Category-agnostic keypoint pose from text descriptions + query image
Input: RGB + text
Summary: CapeLLM replaces support images with annotated keypoints by a query image and detailed text descriptions. Task-specific instructions, dynamic round training, and flexible decoding estimate category-agnostic keypoint pose and spatial uncertainty, evaluated on MP-100.
Scope & Sources
Inclusion: officially accepted ICCV 2025 main-conference papers whose titles or abstracts directly address object 6D / 3D pose, novel-object pose, category-level pose, category-agnostic keypoint pose, or object-pose robustness, confidence, and precision measurement.
Excluded: human pose, relative camera pose and visual localization, scene / camera reconstruction alone, detection or classification alone, workshop papers, and work treating object pose only as a downstream variable or background concept. Ultra-Precision 6DoF and CapeLLM are retained but have different task boundaries from standard BOP-style instance-level 6D pose.
Primary sources: Official ICCV 2025 accepted-paper directory, ICCV 2025 Open Access and each paper's official page. Summaries, classifications, recommendations, and trends are editorial work and do not replace the original papers.
