更新记录
2026 年 6 月 2 日
BOP Writer 改用 pyrender 多进程生成真值掩码和场景元数据。源码提交
2026 年 4 月 6 日
安装文档新增了最小推理环境、首次运行耗时说明、故障排查和可选的 requirements-inference.txt。ONNX 脚本可能自动安装依赖,建议在独立的 venv 或 conda 环境中运行。
2026 年 4 月 4 日
- 新增 MegaPose 位姿精修接口及 RGB-D 示例。
- HccePose 支持通过
hccepose_acceleration选择 ONNX 或 TensorRT;time_dict提供分阶段计时。 - 发布 RGB-D 示例帧
000000至000003。每帧含{stem}_rgb.png、{stem}_depth.png和{stem}_camK.json。参见示例文件和下载说明。 - OpenCV 返回的 BGR 图像直接传给
Tester.predict;MegaPose 调试图使用 BGR 保存。
2025 年 10 月 28 日
更新了标签制备流程。PBR 渲染入口可以自动生成默认的 camera.json。渲染脚本
2025 年 10 月 27 日
发布 cc0textures-512。这套材质库约 600 MB,可在渲染示例中替代约 44 GB 的原版 CC0Textures。
历史路线图
2025 年的路线图以当年年底为目标。当时列为已完成的有七个 BOP 核心数据集的权重、BOP 挑战评测流程,以及物体预处理、数据渲染、YOLOv11 标签与训练、HccePose(BF) 标签与训练的教程。
当时仍在计划中的三项工作是帧间位姿跟踪、用 HccePose(BF) 制备真实场景位姿数据集,以及联合使用 PBR 和真实数据训练。仓库里的视频示例展示的是逐帧独立推理。
Changelog
June 2, 2026
The BOP Writer now uses pyrender with multiple processes to generate ground-truth masks and scene metadata. Source commit
April 6, 2026
The installation guide added a minimal inference setup, first-run timing notes, troubleshooting and an optional requirements-inference.txt. The ONNX scripts may install dependencies automatically. Run them in a dedicated venv or conda environment.
April 4, 2026
- Added the MegaPose refinement interface and RGB-D examples.
- HccePose supports ONNX or TensorRT through
hccepose_acceleration;time_dictprovides stage timings. - Released RGB-D frames
000000through000003. Each frame includes{stem}_rgb.png,{stem}_depth.pngand{stem}_camK.json. See the sample files and download instructions. - Pass OpenCV BGR images directly to
Tester.predict; MegaPose debug images are saved in BGR.
October 28, 2025
Updated label generation. The PBR rendering script can create a default camera.json automatically. Rendering script
October 27, 2025
Released cc0textures-512, a texture library of about 600 MB that can replace the original 44 GB CC0Textures library in the rendering example.
Historical roadmap
The 2025 roadmap targeted the end of that year. It marked three items complete: weights for the seven core BOP datasets, the BOP Challenge evaluation workflow, and tutorials covering object preparation, rendering, YOLOv11 labels and training, and HccePose(BF) labels and training.
Three items were still planned at the time: inter-frame pose tracking, real-scene pose dataset preparation with HccePose(BF), and training on a mix of PBR and real data. The video example in this repository runs inference independently on each frame.