Setting: Rigid objects unseen during training; recovers 6D pose from local model-to-scene RGB-D correspondences

Input: RGB-D + object model

Summary: Learns rotation-invariant geometric descriptors through a Fuse–Describe–Match pipeline, then fuses RGB texture into the descriptor space using attention to reduce ambiguity from symmetry or repeated structures in geometry-only matching. New objects require neither retraining nor large-scale template rendering.

Setting: Rigid objects unseen during training; unifies CAD-based and model-free settings with a few reference images

Input: RGB-D + CAD or posed references

Summary: Renders the object directly when CAD is available; otherwise fits an object-centric neural field from a few posed reference images. Both modes share RefineNet and ScoreNet for immediate pose estimation and tracking of objects unseen during training.

Setting: Single-image RGB instance-level 6DoF pose with known CAD; no iterative refinement

Input: RGB + CAD render

Summary: First classifies a coarse pose and renders the CAD model, then uses multiscale residual correlation to establish low- and high-level correspondences between the input and rendering before regressing fine residuals. Soft probabilistic pose labels handle symmetric objects, with emphasis on a single forward pass.

Setting: Single-frame RGB-D instance-level pose with a known object model; targets real-time correspondence-based estimation without rendering-based refinement

Input: RGB-D + CAD model

Summary: Encodes the CAD surface as hierarchical binary regions. Starting from coarse point-to-surface correspondences, it progressively narrows regions and rejects outliers to obtain 3D–3D correspondences and pose, aiming for high accuracy without expensive rendering-based refinement.

Setting: Given a novel object's CAD model and an initial pose, iteratively refines unseen-object poses using a general optical-flow network

Input: RGB or RGB-D + CAD render

Summary: Predicts dense flow, confidence, and pose updates between the observation and current CAD rendering, using 3D shape constraints and differentiable PnP to maintain projection consistency. Cascaded multiscale correlation volumes support recurrent coarse-to-fine refinement.

Setting: Single RGB image, target CAD, and object detection; fast coarse pose for objects unseen during training

Input: RGB + CAD templates

Summary: Retrieves out-of-plane rotation from a CAD template library with only two rotational degrees of freedom, then uses one local correspondence to recover in-plane rotation, scale, and image translation. Substantially faster than earlier template methods and readily combined with refinement.

Setting: Given target CAD, finds all target instances in cluttered RGB-D and estimates poses for objects unseen during training

Input: RGB-D + CAD templates

Summary: ISM uses SAM / FastSAM to generate open proposals and filters them by semantic, appearance, and geometric scores. PEM treats pose as partial-to-partial 3D matching, using background tokens and coarse-to-fine correspondences to recover SE(3).

Setting: Uses intermediate diffusion-model features to improve image-based pose representations for unseen objects

Input: RGB + object model / templates

Summary: Analyzes viewpoint and object generalization in intermediate features of generative models such as Stable Diffusion, and develops three aggregation architectures at different granularities for object pose. Reports unseen-object improvements mainly on LM, O-LM, and T-LESS.

Setting: Given only one reference RGB image of a novel object, predicts its relative viewpoint in a query image

Input: reference RGB + query RGB

Summary: A pose-conditioned U-Net directly synthesizes discriminative embeddings for candidate viewpoints around the object, then matches the query image to these embeddings. Requires no CAD, object-specific training, or multiview scan, and focuses on relative rotation.

Setting: Reference–query RGB image pair; recovers unseen-object relative pose in one forward pass without CAD

Input: reference RGB + query RGB

Summary: Lifts reference and query images into separate object-centric 3D voxel representations, establishes noise-robust correspondences using weighted closest voxels, and directly solves relative pose by weighted least squares, avoiding exhaustive scoring of large discrete pose-hypothesis sets.

Setting: Uses text to specify a target, localizes it in RGB-D images of two different scenes, and estimates relative 6D pose

Input: text + reference/query RGB-D

Summary: Text prompts guide a VLM to segment the same target in two RGB-D scenes. Prompt-level object semantics are fused with local image features to establish cross-scene point correspondences and solve relative SE(3). The new benchmark builds on REAL275 and Toyota-Light.

Setting: Back-end refinement for category-level RGB-D pose; handles shape differences between observed instances and category shape priors

Input: point cloud + initial pose + shape prior

Summary: Uses HS-layer, learnable affine transformations, and cross-cloud transformation to align observed point clouds with category priors and refine rotation under within-category deformation. Shape priors also support translation and size-error regression, addressing the limited transferability of instance-level refiners.

Setting: Builds a category canonical frame and learns single-image 3D orientation using only unlabeled object-centric videos

Input: unlabeled RGB videos → single RGB

Summary: First aligns canonical cameras across videos using coarse meshes, DINOv2 appearance features, and cyclic distance. The alignments and reconstructed templates then train a single-image dense-correspondence network without manual pose labels, CAD, or RGB-D.

Setting: Learns a conditional pose distribution from a single RGB image, retaining multiple solutions from symmetry and occlusion

Input: RGB

Summary: Trains conditional score-based diffusion on the SE(3) group, using a surrogate Stein score to improve denoising convergence and efficiency. Outputs posterior samples covering multiple equivalent or approximate poses instead of one deterministic solution.

Setting: RGB-D category-level pose and size without compressing instance geometry into a single mean shape

Input: RGB-D / point cloud

Summary: Aligns DINOv2 category semantics with local-to-global SE(3)-invariant geometric features pointwise, forming a dual-stream object representation consistent under coordinate transformations before mapping camera space to canonical space. Evaluated on REAL275 and HouseCat6D.

Setting: Single RGB image with a known object model; recovers instance-level 6D pose through diffusion-based 2D keypoint denoising

Input: RGB + 3D keypoints

Summary: Formulates 2D keypoint detection as conditional reverse diffusion and uses Mixture-of-Cauchy forward noise to match heavy-tailed errors from occlusion and backgrounds. Denoised 2D–3D correspondences are passed to a geometric pose solver.

Setting: Known category, unseen instance; 6D pose and size without a category-specific mean shape

Input: RGB-D / point cloud

Summary: Instance-Adaptive Keypoint Detection selects sparse keypoints that express each instance's geometry. Geometric-Aware Feature Aggregation combines local and global structure to establish keypoint-level correspondences to canonical coordinates.

Setting: Jointly tracks an unknown object's 6DoF pose and currently visible shape in a continuous point-cloud stream

Input: point cloud stream + language

Summary: Models spatiotemporal changes between adjacent point clouds using a pairwise implicit 3D representation. Multilevel 3D–language contrastive alignment enables zero-shot semantic generalization, while the same network outputs per-frame pose tracking and object shape reconstruction.

Setting: Given a model point cloud, discovers an unknown number of instances in a scene point cloud and recovers each instance's rigid transformation

Input: model point cloud + scene point cloud

Summary: At the coarse level, iteratively learns instance-aware superpoint features and instance masks. At the fine level, generates masked correspondence candidates for each instance, then filters and refines them into multiple SE(3) estimates without expensive multi-model fitting.

Setting: Category-level pose, size, and grasp evaluation in real household scenes with challenging materials and large viewpoint changes

Input: RGB + depth + polarization

Summary: Releases a real dataset with synchronized RGB, active stereo depth, and polarization images, covering textureless, reflective, and translucent objects. Pose annotation error is approximately 1.35–1.74 mm, with dense 6D parallel-jaw grasp annotations and category-level baselines.

Setting: Jointly evaluates cameras, exposure / lighting, and 6DoF pose algorithms in industrial workspaces

Input: 13 calibrated multimodal cameras

Summary: IPD contains 2,300 physical scenes, 20 industrial parts, and over 100K object views. Each scene is captured by 13 calibrated cameras at 4 exposures and under 3 lighting conditions. Robot consistency is introduced for scalable high-precision evaluation.

Setting: Matches moved objects across sparse scans over time, estimates each object's 6DoF motion, and accumulates reconstructions

Input: multi-temporal 3D scans

Summary: An SE(3)-equivariant encoder–decoder supports instance matching, registration, and reconstruction across scans. Joint optimization transforms point clouds of the same object at different times into a common frame, recovering object motion and progressively improving geometric completeness.

Setting: Recovers shape, parts, and joint transformations from RGB-D scans of an unknown articulated object in two articulation states

Input: two RGB-D scans

Summary: First reconstructs an object-level implicit shape for each state, then combines images, 3D reconstruction, point correspondences, and kinematic cues to recover part segmentation and the articulation connecting the two states. Supports multiple movable parts without priors on object shape or structure.