3DV 2025 · Paper Collection
Strict 6D object-pose work at 3DV 2025 is concentrated, but the broader scope extends in two directions: building new-object representations from a few unknown viewpoints, and continuously recovering object motion within dynamic scenes and hand–object interaction.
6
Papers in this collection
3
Direct pose / view alignment
3
Dynamic and interacting objects
0
Included Orals
The official 3DV directory lists 142 main-conference papers, including 21 Orals. This collection first includes explicit object 6DoF pose, view alignment, and object-pose outputs, then adds rigid-motion tracking and separate object reconstruction under interaction occlusion. Camera pose alone, generic SfM, and static reconstruction alone are excluded.
Collection in One Sentence
GS-Pose and iFusion use 3D Gaussian Splatting and diffusion priors, respectively, to address pose initialization, unknown-view registration, and low-cost reconstruction for novel-object onboarding, turning prerequisites for object-specific pose into learned representation problems.
SMORE, EgoGaussian, and InterTrack recover pose within dynamic processes where cameras move and hands occlude objects, requiring stable trajectories and reusable object representations. Betsu-Betsu supplies the basis for separating geometry during close interaction.
Research Trends
These themes summarize the 6 included papers by output variables and problem conditions; they do not represent the entire 3DV research distribution.
01 · Object Onboarding
GS-Pose builds an object library from posed RGB views for retrieval and 3DGS refinement; iFusion treats unknown view pose as a latent variable inferred through a diffusion prior. Both reduce acquisition requirements for new objects entering pose pipelines.
02 · Dynamic Object State
SMORE jointly optimizes neural surfaces, ego poses, and object poses in dynamic LiDAR scenes. EgoGaussian separates dynamic objects from backgrounds in egocentric RGB and tracks rigid motion. Neither treats moving objects simply as SLAM outliers.
03 · Interaction & Occlusion
InterTrack uses temporal information to predict smooth object rotation in template-free hand–object video; Betsu-Betsu decodes closely interacting objects as independent SDFs. Both provide more stable object-state foundations for downstream manipulation.
Data Statistics
Each paper belongs to one group by its main output variables; tags retain cross-area attributes.
Direct novel-object 6DoF pose, relative pose between sparse views, or object-pose optimization in dynamic scenes.
Recovers temporally continuous rigid-object motion in egocentric video under camera motion and occlusion.
Separates object geometry under close interaction and mutual occlusion, providing a basis for object coordinate frames and pose modeling.
Interpreting the statistics: Of 6 papers, 3 explicitly output pose or view alignment, but only GS-Pose follows the standard novel-object 6D pose setting. Other work concerns dynamic object-state recovery or interaction reconstruction and should not be compared directly through BOP-style 6D pose metrics.
Recommended Reading
Recommendations reflect problem coverage and methodological representativeness. All included papers are Posters; presentation type does not substitute for editorial assessment.
The collection's only paper directly targeting novel-object 6D pose, combining segmentation, retrieval initialization, and 3DGS render-and-compare refinement into complete object onboarding.
Although distinct from BOP object pose, it recasts unknown-view registration as generative-prior inversion, relevant to low-cost object onboarding.
Places ego and object poses in one global optimization, representing a transition toward reconstructable scene-level representations of multiple moving objects.
Uses only egocentric RGB to address dynamic-object segmentation, scene representation, and rigid-motion tracking together, relevant to wearable perception and interactive robotics.
Combines template-free single-frame reconstruction and temporal object-rotation prediction to address tracking failures under severe hand–object occlusion.
Paper List
Each card gives task boundaries, input modalities, tags, and a short summary. The main link opens the site's coarse and deep reading notes.
3 papers
Direct novel-object 6DoF pose, relative pose between sparse views, or object-pose optimization in dynamic scenes.
Dingding Cai · Janne Heikkilä · Esa Rahtu
Setting: After collecting posed RGB views of an unseen object, localizes it and estimates 6DoF pose in a query image
Input: posed RGB object views + query RGB
Summary: GS-Pose builds representations for three stages: target localization, initial 6D pose retrieval, and refinement using fast differentiable 3D Gaussian Splatting rendering. Reports strong results on LINEMOD and OnePose-LowTexture.
Chin-Hsuan Wu · Yen-Chun Chen · Bolivar Enrique Solarte · Lu Yuan · Min Sun
Setting: Joint view-pose estimation and object reconstruction from two object images with unknown camera poses
Input: two RGB object views with unknown camera poses
Summary: iFusion first inverts a pretrained novel-view diffusion model to estimate relative input-view poses, then fine-tunes an object-specific generator using these views before applying a reconstruction method to sparse-view object reconstruction.
Nathaniel Eliot Chodosh · Anish Madan · Simon Lucey · Deva Ramanan
Setting: Joint background mapping, rigid-object surface reconstruction, and object-pose recovery in large dynamic LiDAR scenes
Input: dynamic LiDAR scans
Summary: SMORE decomposes dynamic environments into background and rigid moving objects, applying global coordinate descent over neural surfaces, ego poses, and object poses. Alternates registration and surface reconstruction, with dynamic-object deskewing for rolling-shutter LiDAR.
2 papers
Recovers temporally continuous rigid-object motion in egocentric video under camera motion and occlusion.
Daiwei Zhang · Gengyan Li · Jiajie Li · Mickaël Bressieux · Otmar Hilliges · Marc Pollefeys · Luc Van Gool · Xi Wang
Setting: Joint scene reconstruction and interacting rigid-object motion tracking from egocentric RGB video
Input: egocentric monocular RGB video
Summary: EgoGaussian automatically separates background and dynamic-object Gaussians, using clip-level online learning to reconstruct scene evolution in temporal order and track interacting rigid objects' 3D motion.
Xianghui Xie · Jan Eric Lenssen · Gerard Pons-Moll
Setting: Video hand–object interaction tracking under hand occlusion without predefined object-shape templates
Input: monocular human-object interaction video
Summary: InterTrack decomposes 4D hand–object tracking into per-frame pose tracking and canonical-shape optimization. Uses temporal information to predict smooth object rotation, trains on 8.3k synthetic interaction sequences, and evaluates on BEHAVE and InterCap.
1 papers
Separates object geometry under close interaction and mutual occlusion, providing a basis for object coordinate frames and pose modeling.
Suhas Gopal · Rishabh Dabral · Vladislav Golyanik · Christian Theobalt
Setting: Separate reconstruction of two closely interacting, heavily mutually occluded objects from multiview RGB
Input: multi-view RGB images
Summary: Betsu-Betsu encodes the interaction scene through a shared multiresolution hash grid, then decodes separate neural SDFs for the two objects. Alpha-blending regularization keeps geometry separate at interaction boundaries, supporting rigid and articulated objects.
Scope & Sources
Inclusion: officially accepted 3DV 2025 main-conference papers directly outputting novel-object 6DoF pose, input-view pose / alignment, object pose, or rigid-object motion in dynamic and interactive settings. One separate-geometry reconstruction paper on close object interaction is retained as a prerequisite representation reference.
Excluded: camera pose alone, SfM, generic point-cloud registration, 3D detection, human pose, static object reconstruction alone, and work treating object pose only as an indirect downstream variable. iFusion estimates input-view alignment and is distinguished from 6D object-pose benchmarks.
Primary sources: Official 3DV 2025 paper directory, the directory's official schedule CSV , and public PDFs. Summaries, classifications, and trends are editorial work and do not replace the original papers.
