项目结构
本项目包含应用推理、公开 WAPR 接入和双语网站源码。训练与渲染的阶段脚本见下方公开上游索引。
HccePose/ RGB 推理、网络与 BOP 加载器
s1_*.py ... s4_*.py 公开准备、训练与评测阶段
Refinement/ WAPR 接入与可选 MegaPose 接口
examples/inference/ predict_image.py、predict_wapr.py、postprocess_wapr.py
docs/zh/ 中文文档
docs/en/ 英文文档
docs/theme/ 共享视觉主题
docs/assets/ 图片与展示素材
pages/ 构建后的静态网站
show_vis/ 公开示例图片
media/ 公开视频预览
tests/ 接口与文档检查
公开阶段脚本
以下文件位于公开上游固定版本根目录;点击可查看源码,参数在脚本中显式设置。
| 脚本 | 用途 |
|---|---|
| s1_p1_obj_rename_center.py | 模型居中与编号 |
| s1_p2_obj_symmetry.py | 旋转对称信息 |
| s1_p3_obj_infos.py | BOP 模型元数据 |
| s2_p1_gen_pbr_data.py | PBR 场景渲染 |
| s3_p1_prepare_yolo_label.py | 检测标签转换 |
| s3_p2_train_yolo.py | 检测器训练 |
| s4_p1_gen_bf_labels.py | 前后表面坐标标签 |
| s4_p2_train_bf_pbr.py | 逐物体位姿训练 |
| s4_p2_test_bf_pbr.py | 裁剪图验证 |
| s4_p2_test_bf_pbr_bop_challenge.py | BOP 预测导出 |
数据与权重目录
网格和元数据位于 <dataset>/models/,检测器位于 <dataset>/yolo11/,位姿权重位于 <dataset>/HccePose/obj_XX/best_score/。训练场景和两套表面标签分别位于 train_pbr/、train_pbr_xyz_GT_front/、train_pbr_xyz_GT_back/。
推理下载布局与相机文件见资源说明。脚本组织不改变 BOP 物体编号、像素坐标和毫米模型单位。
构建网站
python -m pip install -r docs/requirements.txt
python docs/build.py
python -m http.server 8768 --directory pages
网站构建无需 GPU,也不运行推理。其他使用说明见文档导航。
Project structure
This project contains application inference, the public WAPR adapter and bilingual website sources. Public upstream training and rendering stage scripts are indexed below.
HccePose/ RGB inference, networks, BOP loader
s1_*.py ... s4_*.py Public preparation, training and evaluation stages
Refinement/ WAPR adapter and optional MegaPose interface
examples/inference/ predict_image.py, predict_wapr.py, postprocess_wapr.py
docs/zh/ Chinese documentation
docs/en/ English documentation
docs/theme/ Shared visual theme
docs/assets/ Figures and presentation assets
pages/ Generated static website
show_vis/ Public example figures
media/ Public video preview
tests/ API and documentation checks
Public stage scripts
These files live in the fixed public upstream root. Follow the links for source; parameters are explicit in each script.
| Script | Purpose |
|---|---|
| s1_p1_obj_rename_center.py | Center and name meshes |
| s1_p2_obj_symmetry.py | Rotational symmetry |
| s1_p3_obj_infos.py | BOP model metadata |
| s2_p1_gen_pbr_data.py | PBR rendering |
| s3_p1_prepare_yolo_label.py | Detection labels |
| s3_p2_train_yolo.py | Detector training |
| s4_p1_gen_bf_labels.py | Front/back coordinate labels |
| s4_p2_train_bf_pbr.py | Per-object pose training |
| s4_p2_test_bf_pbr.py | Crop evaluation |
| s4_p2_test_bf_pbr_bop_challenge.py | BOP prediction export |
Data and checkpoint layout
Meshes and metadata live in <dataset>/models/, the detector under <dataset>/yolo11/, and pose checkpoints under <dataset>/HccePose/obj_XX/best_score/. Training scenes and surface labels use train_pbr/, train_pbr_xyz_GT_front/, and train_pbr_xyz_GT_back/.
See Assets for download layouts and camera files. Script organization does not change BOP object IDs, pixel coordinates or millimeter model units.
Build the website
python -m pip install -r docs/requirements.txt
python docs/build.py
python -m http.server 8768 --directory pages
Building the site needs no GPU and runs no inference. See the documentation index for other usage guides.