Setting: Joint category-level rotation, translation, and scale; precise task boundaries await the official PDF

Input: Awaiting verification from the official PDF

Summary: The official title explicitly applies geodesic flow matching and geometric priors to category-level 9DoF pose. The ECCV page currently provides no abstract or full text, so input modalities, architecture, datasets, and numerical results are not inferred.

Setting: Category-agnostic pose and occlusion robustness; output protocol awaits the official PDF

Input: Awaiting verification from the official PDF

Summary: The official title indicates conditional flow matching for occlusion-resilient category-agnostic pose estimation. Without a public abstract or PDF, keypoint definitions, support-image settings, and evaluation numbers are not inferred from the title.

Setting: Consistent semantic orientation for large-scale 3D assets and zero-shot point-cloud orientation estimation

Input: 3D object assets / point clouds

Summary: CanoVerse builds a canonical-orientation dataset with 1,156 categories and approximately 320K objects. Compact candidate generation and lightweight human discrimination reduce per-object alignment from minutes to seconds. Alongside generation and retrieval, it evaluates zero-shot orientation estimation for out-of-distribution point clouds.

Setting: Historical instance-structure memory for category-level pose of current unseen instances

Input: RGB-D / point cloud (according to the preprint)

Summary: MemPose stores structural representations of past instances in a dynamically updated external memory buffer, beyond fixed parameters or static shape priors. It evaluates cross-instance gains on REAL275, CAMERA25, HouseCat6D, and Wild6D.

Setting: Part pose and shape hypotheses for category-level articulated objects; precise outputs await the official PDF

Input: Awaiting verification from the official PDF

Summary: The official title indicates pose–shape hypothesis generation and verification for category-level articulated objects. Hypothesis spaces, joint parameterization, input modalities, and experimental conclusions remain unverified until full text is available.

Setting: Reference-prior-guided category-level deformation and pose alignment; details await the official PDF

Input: Awaiting verification from the official PDF

Summary: The official title confirms that PriorPose combines reference-guided deformation and alignment. Without an abstract on the official page, the reference format, CAD / depth requirements, and main results are not inferred.

Setting: Large-vocabulary pose estimation; task dimensions and protocol await the official PDF

Input: Awaiting verification from the official PDF

Summary: The title describes pose estimation over extensive classes using rich structure prototypes. Available information does not establish whether outputs are 3D rotation, full 6D/9D pose, or category-agnostic keypoints; inclusion boundaries will be reviewed when the paper is available.

Setting: Unseen-object 6D pose from one or a few unposed reference views

Input: reference RGB(/D) + query RGB(/D)

Summary: PANY learns cross-view-consistent structure and alignment cues through a multiview Transformer geometry backbone and aggregates optional unposed auxiliary views through pose-graph canonical registration. The preprint reports gains of 12 percentage points on YCB-V and over 20 percentage points on LM-O.

Setting: Monocular-depth cues for 6D pose; object setting awaits the official PDF

Input: Monocular RGB / predicted depth (unverified)

Summary: The official title confirms monocular-depth cues for 6D object pose estimation. Before an abstract or full text is available, the source of depth—frozen foundation models or joint training—and benchmark results remain unspecified.

Setting: Zero-shot 6D object pose aiming to avoid refinement

Input: Awaiting verification from the official PDF

Summary: The official title specifies unified flow-based estimation, zero-shot inference, and no refinement. Public evidence does not yet establish flow type, object priors, input modalities, or speed–accuracy tradeoffs, so no numerical results are listed.

Setting: Unseen-object pose refinement under large initial angular errors

Input: Awaiting verification from the official PDF

Summary: The ECCV full text is not yet public. BOP method records describe WAPR as wide-angle refinement handling initial errors up to approximately ±90°, using multiple candidates and iterative updates for unseen-object pose. Training, inputs, and official paper results remain to be verified.

Setting: Industrial-bin 6D pose benchmark with reflective, symmetric, textureless, and heavily occluded objects

Input: multi-sensor RGB / grayscale / depth

Summary: The ECCV accepted version provides 75 multiview real scenes and approximately 273K annotations of metallic and highly symmetric objects. Antireflective spray, multiview depth fusion, and semiautomatic review produce millimeter-level annotations. Several state-of-the-art methods degrade substantially relative to household benchmarks.

Setting: Geometrically consistent image generation for 6D object pose training

Input: Awaiting verification from the official PDF

Summary: The title explicitly constrains generated images to preserve geometry required for 6D pose, directly supporting pose training. The generator, conditioning signals, annotation-preservation mechanism, and downstream gains await verification from the official full text.

Setting: Simultaneous tracking and online reconstruction of multiple unknown rigid objects in monocular RGB-D video

Input: RGB-D video + sparse point prompts

Summary: Point2Pose initializes from sparse image points on objects, obtains temporal correspondences with a long-range 2D point tracker, and combines depth with an object-centric 3D keypoint map to recover pose while incrementally reconstructing a TSDF. The preprint reports performance comparable to state of the art on heavy-occlusion benchmarks, with additional multi-object support and complete-occlusion recovery.

Setting: Joint hand–object 3D pose from one RGB image with unseen objects and novel interactions

Input: RGB + predicted point cloud + text descriptions

Summary: GenHOI encodes objects, hands, and interactions with hierarchical semantic prompts and uses masked modeling over images, text, and internal point clouds. Predicted hand rotation / articulation then constrains object 6D pose. Reported object ADD-0.5D and ADD-S on DexYCB S3 are substantially higher, while hand metrics are only close to UniHOPE; 18 FPS excludes offline text generation.

Setting: Joint two-hand and object pose, mesh, and contact recovery from egocentric in-the-wild RGB

Input: egocentric monocular RGB

Summary: Introduces EPIC-Contact with 2,272 clips and 62.3K frames. HOPformer injects WiLoR hand priors into object features through 12 cross-attention layers. Against the strongest JointTransformer, EPIC SR@0.05 rises from 17.6% to 29.8% and CDev drops from 30.1 to 20.7 mm; the abstract's near-doubling / 75% reduction more closely reflects comparison with ArcticNet-SF. The main pose table uses oracle meshes; automatic classification accuracy is 52.9%.