WAPR model weights / WAPR 模型权重

WAPR refines the 6D pose of an object absent from pose-model training. Provide RGB-D, camera intrinsics, a metric mesh, and a mask or box. Training covers candidate rotation errors up to 90°. The default recipe uses 12 hypotheses, three WAPR updates, two SAPR updates, and WBPS ranking. A metric mesh is still required at inference.

WAPR 修正位姿模型训练时未见过的物体的 6D 位姿。输入为 RGB-D、相机内参、米制网格，以及 mask 或框。训练覆盖高达 90° 的候选旋转误差。默认流程使用 12 个候选、三次 WAPR 更新、两次 SAPR 更新，再由 WBPS 排序。推理时仍需提供米制网格。

Project page / 项目页 · Source / 源码 · Documentation / 文档

Capabilities / 功能

Mask or box inputs, batched RGB-D pose estimation, wide-angle refinement and candidate scoring. Tracking examples include large inter-frame motion and optional lost-track recovery. Optional reconstruction can supply a mesh for pose estimation; pose inference still requires that mesh. See the tracking comparisons and batch measurement conditions.

支持 mask 或框输入、RGB-D 批量位姿估计、广角修正及候选评分。跟踪示例涵盖大帧间运动和可选的跟踪丢失补救。可选重建流程可提供网格；位姿推理仍需该网格。详见跟踪对照与批量测量条件。

Models / 模型

File / 文件	Purpose / 用途	Input / 输入
wapr_w_mask.pth	Wide-angle refinement with a mask cue / 带掩码提示的广角位姿修正	RGB, mask, XYZ — 7 channels / 7 通道
wapr_wo_mask.pth	Wide-angle refinement without a mask cue / 无掩码的广角位姿修正	RGB, XYZ — 6 channels / 6 通道
sapr.pth	Small-angle refinement after WAPR / WAPR 之后的小角位姿修正	RGB, XYZ — 6 channels / 6 通道
wbps.pth	Candidate-pose selection and group scoring / 候选位姿选择与组评分	RGB, XYZ — 6 channels / 6 通道

These four models were trained on the SA6D synthetic RGB-D dataset. An unseen object still requires its metric mesh at inference. The default source configuration uses twelve initial pose hypotheses, three WAPR updates, two SAPR updates and WBPS selection. See the documentation for model inputs, symmetry and score interpretation.

这四份模型均基于 SA6D 合成 RGB-D 数据集训练。未见物体在推理时仍需提供米制网格。源码默认配置使用十二个初始候选姿态、三次 WAPR 更新、两次 SAPR 更新和 WBPS 选择；输入要求、对称性及评分含义见文档。

Usage / 用法

Use a compatible Linux NVIDIA GPU environment with CUDA-enabled PyTorch. Install WAPR and prepare the core environment:

使用兼容的 Linux NVIDIA GPU 环境及支持 CUDA 的 PyTorch，安装 WAPR 并准备核心环境：

python -m pip install -U wapr==0.0.3
python -m wapr.bootstrap
python -c "from wapr.bootstrap import export_examples; export_examples('wapr_examples')"
python wapr_examples/02_one_category_one_instance.py

Weights and example inputs are obtained when needed. See the installation guide for prerequisites, resource locations and optional features. Optional gated models require your own access or local files. Depth and mesh coordinates are in meters; camera intrinsics are in pixels.

权重与示例输入按需获取。系统前提、资源位置及可选功能见安装引导。受控模型需要用户自己的访问权限或本地文件。深度和网格坐标单位为米，相机内参单位为像素。

Example inputs / 示例输入

The samples/pose_lmo/ and samples/bop/ directories contain small RGB-D and mesh excerpts for the examples. PROVENANCE.txt files describe the original sources and units. Dataset licenses apply independently of the model-weight license. See the dataset descriptions for their tasks and challenges.

samples/pose_lmo/ 与 samples/bop/ 提供示例所用的少量 RGB-D 和网格摘录。PROVENANCE.txt 说明原始来源与单位；数据集条款独立于模型权重许可。数据用途与难点见数据集说明。

License / 许可

The four first-party model weights use CC BY-ND 4.0; see WEIGHTS_LICENSE.txt and CC-BY-ND-4.0.txt. Commercial use is free. Original redistribution requires attribution; adapted weights may not be shared under this license. For enterprise versions or customization, contact Yulin Wang: shopedataset@gmail.com. Source code uses LGPL-2.1-only. This model-card license does not cover datasets or third-party resources.

四份第一方模型权重采用 CC BY-ND 4.0，见 WEIGHTS_LICENSE.txt 与 CC-BY-ND-4.0.txt。免费允许商用；原样再发行须署名，本许可不允许共享修改版权重。企业版本与定制需求联系 Yulin Wang：shopedataset@gmail.com。源码采用 LGPL-2.1-only；模型卡许可不覆盖数据集与第三方资源。