WAPR, wide-angle pose refinement
WAPR

Example overview示例总览

Complete Installation and check input formats in Data and examples. Example 01 introduces 2D detection and segmentation; 02 estimates a single object's pose from a supplied region; 03 uses published detection boxes.

先完成安装，并在数据与示例核对输入格式。示例 01 介绍 2D 检测与分割，02 使用给定区域估计单个物体的位姿，03 接入已公布的检测框。

The 17 public examples follow the documentation's progression: 01–08 cover detection, pose interfaces, custom data, BOP export and tracking; 09 adds TACO[5] multi-object tracking; 10 is the ROBI[7] experiment; 11–12 cover reconstruction and cross-scene pose; 13–17 cover robot simulation. Edit the visible recipe variables for your inputs. Supporting files stay in the folder bearing the entry script's name.

全部 17 个公共示例按文档的使用逻辑连续编号：01–08 为检测、位姿接口、自定义数据、BOP 导出与跟踪；09 为 TACO[5] 多物体跟踪；10 为 ROBI[7] 实验；11–12 为重建与跨场景位姿；13–17 为机器人仿真。通过脚本中的可见配置变量设置输入，辅助文件放在与入口同名的目录中。

Script脚本 Learning objective学习目标 Page页面
2D detection and 6D pose interfaces2D 检测与 6D 位姿接口
examples/01_one_rgb_detect_segment.py One LM-O[4] RGB frame: build a CAD template bank, then return 2D boxes and masks. No pose step.一帧 LM-O[4] RGB：建立 CAD 模板库，返回 2D 框和 mask；此步不估计位姿。 2D detector call2D 检测调用
examples/02_one_category_one_instance.py Estimate the pose of one known object with estimate_one_category_one_instance; see the result on the pose page.用 estimate_one_category_one_instance 估计单个已知物体的位姿；结果图见位姿页。 One category, one instance单类别、单实例
examples/03_published_boxes_to_pose.py Read the published detection excerpt for image 307 and estimate all retained instances in one batch, with independent hypothesis groups.读取第 307 帧的已公布检测摘录，同帧保留实例批量估计，各实例的候选组独立评分。 Published boxes已公布的检测框
examples/04_6d_localization.py 6D localization. Many known categories, each with a fixed count. LM-O scene 2, image 307: eight categories, one each.6D 定位。多个已知类别，每个类别数量固定。LM-O 场景 2 第 307 帧：八个类别，每个 1 个。 Localization定位
examples/05_bop_6d_detection.py BOP challenge, 6D detection. Many categories, many instances, count unknown: 2D, then estimate_frame_many_categories_many_instances.BOP 挑战，6D 检测。多类别、多实例，数量事先不知道：先 2D，再 estimate_frame_many_categories_many_instances。 Detection检测
examples/06_custom_scene.py Load a custom RGB-D folder and CAD meshes; the documentation demonstrates a real wedge capture.读取自定义 RGB-D 目录与 CAD 网格；文档展示三角块实拍案例。 Custom data自定义数据
examples/07_write_bop_pose_csv.py Export a BOP pose CSV from live 2D detections or a supplied JSON. Both paths batch each frame's retained instances; the JSON path filters input masks across categories first.将现场 2D 检测或外部 JSON 候选导出为 BOP 位姿 CSV。两条路径均按帧批量估计；JSON 路径先对输入掩码做跨类别抑制。 Result export结果导出
examples/08_ycbineoat_one_instance.py Initialize one YCBInEOAT[6] track from a predicted mask, then track its target mask and update its pose. Recovery compares tracked and re-estimated poses in a supported two-candidate group.从预测掩码初始化一条 YCBInEOAT[6] 轨迹，再跟踪目标掩码并更新位姿；恢复分支将跟踪与重估位姿组成受支持的两候选评分组。 One instance tracking单实例跟踪
Application cases应用案例
examples/09_taco_many_instances.py Predict SAM 2[2] masks from saved RGB clicks; share DINOv2[1] frame features, batch both tracks and recovery candidates, and preserve each track's reliable state.保存的 RGB 点击点预测 SAM 2[2] 掩码；共享 DINOv2[1] 帧特征，批量更新双物体及补救候选，各轨迹保留可靠状态。 Multi-object tracking多物体跟踪
examples/10_robi_zigzag.py Detect and estimate multiple instances of one CAD in ROBI[7] Zigzag, scene 4, view 0. Pose inference is batched; the Zigzag viewer describes its candidate and iteration settings in the ROBI case.在 ROBI[7] Zigzag 场景 4 第 0 帧检测同一 CAD 的多个实例并批量估计位姿；Zigzag 查看器的候选与迭代设置见 ROBI 案例。 ROBI
examples/11_reconstruct_object.py One RGB-D and a point or sentence: reconstruct, size, select a pose, update shape with RoMa[3], and save a matched mesh/pose bundle.一张 RGB-D 与点或句子：重建、定标、选择位姿、RoMa[3] 更新形状，保存配套网格与位姿。 3D generation3D 生成
examples/12_cross_scene_pose.py Read the saved reconstruction, detect its top-scoring instance in another RGB-D image, then estimate its pose.读取保存的重建网格，在另一张 RGB-D 中选择最高分检测实例，再估计位姿。 Cross-scene pose estimation跨场景位姿估计
examples/13_known_mesh_place.py Panda with table and wrist RGB-D: estimate the known bottle and box, grasp, measure the held-object offset, and compare planned and corrected placement.Panda 桌面与腕部 RGB-D：估计已知瓶子和盒子，抓取、测量夹持偏差，对照计划与修正放置。 Robot机器人
examples/14_bridge_tasks.py WidowX uses the first-frame estimated pose to approach, close, lift, transfer, release and retreat in Bridge[8] tasks.WidowX 在 Bridge[8] 任务中根据首帧估计位姿接近、合爪、抬升、搬运、松爪与撤回。 Robot机器人
examples/15_xarm_cube.py xArm6/Robotiq grasps a cube at its estimated world pose and releases it at the marker; each planned step records RGB-D.xArm6/Robotiq 按估计的世界位姿抓取方块，在标记处松爪；每个规划步记录 RGB-D。 Robot机器人
examples/16_follow_saved.py Track saved robot RGB-D: initialize from verified estimates, share image features, batch all same-frame pose updates, and report errors afterward.跟踪保存的机器人 RGB-D：用已核验估计初始化、共享图像特征、批量更新同帧位姿，并事后报告误差。 Robot机器人
examples/17_virtual_grasp.py Convert saved estimated poses to world coordinates, construct virtual gripper axes, and project the grasp into RGB. Physical control is shown in 13–15.将保存的估计位姿转到世界坐标，构造虚拟夹爪轴并投影到 RGB；实际控制流程见 13–15。 Robot机器人

References and licenses参考文献与许可

  1. DINOv2 — Apache-2.0. Code and official DINOv2 weights; retain copyright and license.源码与官方 DINOv2 权重；保留版权和许可。 · GitHubGitHub · License/notice 1许可/声明 1
    Oquab et al. DINOv2: Learning Robust Visual Features without Supervision. arXiv:2304.07193, 2023. · Paper论文 ↩ ↩
  2. SAM 2 — Apache-2.0; cctorch: BSD-3-Clause. Native source and official checkpoints; cctorch carries an additional BSD notice. Ultralytics-converted files require checking their distributor terms.原生源码与官方权重；cctorch 另附 BSD 声明。Ultralytics 转换文件还需核对分发方条款。 · GitHubGitHub · License/notice 1许可/声明 1 · License/notice 2许可/声明 2
    Ravi et al. SAM 2: Segment Anything in Images and Videos. arXiv:2408.00714, 2024. · Paper论文 ↩ ↩
  3. RoMa — MIT. Matching source; separately obtained checkpoints and dependencies retain upstream terms.匹配源码；另行获取的权重与依赖保留上游条款。 · GitHubGitHub · License/notice 1许可/声明 1
    Edstedt et al. RoMa: Robust Dense Feature Matching. CVPR 2024. · Paper论文 ↩ ↩
  4. LM-O data — CC-BY-SA-4.0. Data and model excerpt, not the loader code.数据与模型摘录，不是读取代码。 · Original source原始来源 · GitHub: BOP toolkitGitHub：BOP 工具集
    Brachmann et al. Learning 6D Object Pose Estimation Using 3D Object Coordinates. ECCV 2014. · Paper论文 ↩ ↩
  5. TACO data — Not separately verified / 未单独核实. Data terms have not been separately verified. The official repository is the original source; the referenced Hugging Face acquisition mirror has no explicit license field. Confirm the data owner's terms for redistribution or commercial use.数据条款尚未单独核实。官方仓库为原始出处；实际获取数据的 Hugging Face 镜像数据卡未声明明确许可字段。再分发或商业使用须核实数据权利方的条款。 · GitHubGitHub
    Liu et al. TACO: Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding. CVPR 2024. · Paper论文 ↩ ↩
  6. YCBInEOAT data — Not separately verified / 未单独核实. Tracking data permission must be checked at the original archive; the comparison code license does not establish the data license.跟踪数据许可须在原始归档核实；对照代码的许可不能作为数据许可。 · GitHubGitHub · GitHubGitHub
    Wen et al. se(3)-TrackNet: Data-driven 6D Pose Tracking by Calibrating Image Residuals in Synthetic Domains. IROS 2020. Introduces YCBInEOAT. · Paper论文 ↩ ↩
  7. ROBI data — Not separately verified / 未单独核实. The saved public poses and dataset are credited to ROBI; an independent grant to redistribute them has not been verified.保存的公开位姿和数据均注明 ROBI 来源；其再分发授权尚未独立核实。 · GitHubGitHub
    Yang et al. ROBI: A Multi-View Dataset for Reflective Objects in Robotic Bin-Picking. IROS 2021. · Paper论文 ↩ ↩ ↩ ↩
  8. BridgeData V2 · Original source原始来源 · GitHubGitHub
    Walke et al. BridgeData V2: A Dataset for Robot Learning at Scale. arXiv:2308.12952, 2023. · Paper论文 ↩ ↩

Model weights and dataset/task assets may have separate terms. WAPR's source license does not replace them.模型权重、数据集与任务资源可能有独立条款；WAPR 源码许可不替代这些许可。