Setting: Instance-level 6DoF pose of new objects from a few reference images without CAD

Input: few-shot reference RGB + query RGB

Summary: Cas6D uses a self-supervised ViT to reduce false target detections with very few references, retrieves top-K pose candidates, then refines them through multiscale warped feature volumes and progressively narrower pose bins. In the 32-shot setting, it reports Proj-5 gains over OnePose++ / Gen6D on LINEMOD and GenMOP, respectively.

Setting: Learns implicit object models for 6D pose from weakly annotated object views without accurate CAD

Input: weakly labeled RGB object views + query RGB

Summary: NeRF-Feat first learns implicit object geometry with NeRF, then jointly trains CNN features through contrastive learning to account for viewpoint invariance and symmetry. Query-image features match 3D NeRF features to solve pose in the neural-field frame, evaluated on LM, LM-Occlusion, and T-LESS.

Setting: Test objects come from unseen categories, with no ground-truth object location in the query

Input: reference templates + query RGB

Summary: LocPoseNet addresses object locations often assumed known in unseen-object pose, introducing distributed reference kernels and multiscale correlation. Decoupled scale-aware / scale-robust features predict translation-related parameters, while synthetic data separately evaluates multiple location-noise conditions.

Setting: Single-forward-pass detection and direct 6D pose regression for multiple known objects from one RGB image

Input: single monocular RGB image

Summary: YOLO-6D-Pose jointly predicts classes, locations, and poses on YOLOX. Pose parameterization, dedicated losses, and 6D augmentation eliminate keypoints, PnP, and refinement postprocessing, shortening multi-object inference through an end-to-end single-stage design.

Setting: Given approximate 3D boxes in scenes scanned by a moving RGB-D camera, jointly retrieves object categories, CAD, and poses

Input: RGB-D scan + CAD database + approximate 3D box

Summary: After comparing CAD–scan similarity measures, HOC-Search selects a depth-and-mask render-and-compare objective. A search tree over categories, models, and poses and a Monte Carlo Tree Search-like algorithm accelerate joint retrieval, with reported speedups of approximately 10×–120× over exhaustive search.

Setting: Establishes local correspondences for relative pose between wide-baseline images of the same object with little shared visibility

Input: object-centric RGB image pairs + estimated 3D signals

Summary: LFM-3D encodes normalized object coordinates or monocular depth as a low-dimensional 3D positional signal for a graph-neural-network matcher. Its output differs from BOP object-to-camera pose, but improved matching substantially benefits relative posing of in-the-wild object-image pairs.

Setting: Joint viewpoint and shape recovery from approximately five images, with unknown object categories and camera poses

Input: five sparse RGB object views

Summary: FORGE predicts per-view 3D features, estimates relative camera poses through cross-view correlation, then fuses aligned features into a neural radiance field. Requires no ground-truth 3D shapes and demonstrates generalization to unseen categories and in-the-wild object images.

Setting: Recovers 6D camera poses for a set of 2–8 object images under symmetry and weak texture

Input: 2–8 sparse RGB object views

Summary: Building on pairwise relative-rotation distributions, RelPose++ jointly reasons over multiple images through attention to resolve pairwise symmetry ambiguities and adds translation prediction decoupled from rotation ambiguity. Outputs can initialize sparse-view object reconstruction.

Setting: Joint hand pose, articulated-object part pose, and contact states in real hand–object images

Input: RGB hand-object observations + simulated interaction priors

Summary: ContactArt collects object-pose and contact annotations in a physics simulator through iPhone-driven visual teleoperation. It learns a part-arrangement discriminator and a contact-region diffusion model, using both priors to constrain hand and articulated-object pose estimation in real images.

Setting: Mutual hand–object occlusion in sparse video degrades interaction reconstruction and pose accuracy

Input: sparse multi-view hand-object videos

Summary: NCRF jointly learns a contact optimization field, a hand–object radiance field, and an observation-to-canonical motion field, mutually constraining contact, pose, and appearance. Evaluates novel-view rendering and pose accuracy on HO3D and DexYCB.

Setting: Tracks rigid-object pose while estimating velocity and friction in sliding and collision scenes

Input: RGB-D video + mesh-based object shapes

Summary: Uses a differentiable physics simulator as the Extended Kalman Filter transition model, jointly filtering object position, orientation, velocity, and friction coefficient under contact and friction constraints. Explicitly enforces physically feasible trajectories beyond visual-only framewise tracking.

Setting: Joint dynamic-scene reconstruction and tracking of all dense scene elements from multiple cameras and times

Input: multi-view, multi-timestep RGB images

Summary: Dynamic 3D Gaussians keeps Gaussian color, opacity, and shape persistent while positions and orientations change over time. Local-rigidity regularization and dynamic novel-view reconstruction yield dense 6DoF trajectories for nonrigid reconstruction and tracking.