Setting: Instance-level 6D pose from synthetic training to real industrial stacked scenes

Input: point cloud / depth

Summary: Builds on pointwise keypoint regression and Hough voting, adding symmetry-class-aware keypoint selection, candidate filtering, and student–teacher geometric pseudo-labels. Reports 96% AP on Siléane and improved generalization on Parametric.

Setting: Rusty, reflective, textureless industrial parts with multiple symmetries

Input: point cloud

Summary: Uses attention-guided multiscale point-cloud features, symmetry-aware rotation loss, and center-distance-sensitive translation loss to regress instance centers and poses pointwise. Two-stage clustering separates instances and produces 6D poses, with reported F1 and Recall gains.

Setting: 6DoF pose of industrial parts with unknown dimensional parameters in dense stacks

Input: point cloud / depth

Summary: Treats parametric shapes as a category, optimizes sparse keypoints and groups for geometric uniqueness, and predicts scene and template keypoints coarse to fine. Reports improved learning and unseen-size generalization on Siléane and Parametric.

Setting: Detection, pose, and online refinement for industrial components spanning different sizes

Input: RGB

Summary: Uses scalable EfficientNet and enhanced feature pyramids for small-object features, combining 2D detection and geometry for adaptive refinement. Unified scaling hyperparameters control backbone and refinement complexity, evaluated on LINEMOD, a photovoltaic-component dataset, and parallel-robot experiments.

Setting: Synthetic-only training with transfer to multiple real category-level pose datasets

Input: RGB-D / multi-modal

Summary: Predicts dense canonical maps through conditional diffusion for partial-shape recovery and correspondence, fusing multimodal input for stronger geometry. Trains only on synthetic data while demonstrating strong sim-to-real transfer across real datasets.

Setting: Category-level pose + size from moving-camera RGB video without high-quality sensor depth

Input: RGB video + optional scale cue

Summary: Estimates camera pose and metric depth through monocular dense SLAM and scale readings, predicts canonical object representations, obtains per-view object poses by registration, then jointly optimizes an object-level pose graph. Gains are stronger with poor or absent depth.

Setting: Zero-shot category-level 6D pose from a few RGB references without depth

Input: RGB

Summary: Synthesizes reference views with Zero-1-to-3, obtains coarse pose by feature matching, and refines through online optimization designed for within-category geometry differences. CO3D experiments test fewer references and depth-free estimation.

Setting: Strawberry detection, 6DoF pose, and dimensions under occlusion and instance-shape variation

Input: RGB

Summary: Automatically generates pose- and size-labeled strawberry data in Ignition Gazebo and trains a single-stage model with domain randomization. Reports simulated-data 3D IoU AP of 84.77%, up to 60 FPS, and transfer to real occluded scenes.

Setting: Few real RGB images without accurate CAD or a manual PBR pipeline

Input: reference RGB + query RGB

Summary: Semiautomatically learns NeuS2 representations through SfM and object-agnostic segmentation, generates training images through novel-view synthesis and cut-and-paste, then trains dense-correspondence SurfEmb. Compares CAD-based and CAD-free methods on LM-O and captured objects.

Setting: 6D pose of a tool's contact-performing functional end when hands occlude the tool body

Input: RGB-D

Summary: Segments tool end-effector affordances from RGB-D, predicts category-level functional-end poses through diffusion, and handles ambiguity with symmetry-aware representations. Releases TOOLEE with affordance segmentations and use-defined poses.

Setting: Rapid physically plausible pose-training scenes assembled from real object and environment scans

Input: RGB / depth / mask / 6DoF labels

Summary: Scans environments and objects as separate 3D Gaussian representations, uses extracted meshes and physics for natural placements, then renders RGB, depth, masks, boxes, and 6DoF poses. Releases an instant-noodle-cup dataset and evaluates sim-to-real through pose networks and real grasping.

Setting: Joint ego and multi-object pose estimation in highly dynamic road scenes

Input: camera + LiDAR / sensor fusion

Summary: Uses a multiprobability-fusion tracker for object association, 3D Kalman velocity priors and optical-flow decoupling to remove dynamic points, then jointly refines camera and object poses through factor constraints. Listed as a dynamic-state extension rather than directly ranked with BOP rigid-instance evaluation.