快速、简单的2D-6D位姿估计：Gen6D算法复现（pytorch 1.12.1 + cu113)

摘要

记录一下复现Gen6D的过程，主要记录其中遇到的一些bug，主要参考了博主牙牙要健康、Fate-Sky-CSDN博客这两位博主的博客。

参考文章：

【6D位姿估计】【深度学习】windows10下Gen6D代码Pytorch实现

6D姿态估计网络【Gen6D复现】CUDA11.3+cuDNN8.2.1-CSDN博客

主要步骤

clone源代码库

配置环境，这个我就不重复了

快速运行

分割视频 改文件夹名字视频名字

python prepare.py --action video2image --input data/custom/video/support_2.mp4 --output data/custom/support_2/images --frame_inter 5 --image_size 960 --transpose

生成点云 改文件加名字时间很长

python prepare.py --action sfm --database_name custom/support_2 --colmap \COLMAP3.8\lib\colmap.exe

分割点云--手动

保存切割点云 object_point_cloud.ply，保存XZ方向文件 meta_info.txt

执行推理 测试视频可以多弄几个

python predict.py --cfg configs/gen6d_pretrain.yaml --database custom/support --video data/custom/video/support2.mp4 --resolution 960 --transpose --output data/custom/support/test --ffmpeg ffmpeg.exe

主要bug1

复制代码

port --colmap colmap.exe
OMP: Error #15: Initializing libiomp5md.dll, but found libiomp5md.dll already initialized.
OMP: Hint This means that multiple copies of the OpenMP runtime have been linked into the program. That is dangerous, since it can degrade performance or cause incorrect results. The best thing to do is to ensure that only a single OpenMP runtime is linked into the process, e.g. by avoiding static linking of the OpenMP runtime in any library. As an unsafe, unsupported, undocumented workaround you can set the environment variable KMP_DUPLICATE_LIB_OK=TRUE to allow the program to continue to execute, but that may cause crashes or silently produce incorrect results. For more information, please see http://www.intel.com/software/products/support/.

关于OMP: Error #15: Initializing libiomp5md.dll, but found libiomp5md.dll already initialized.错误解决方法 - 知乎 (zhihu.com)

删除了当前环境内 bass_gen 文件夹的 C:\Users\StealMoon.conda\envs\bass_gen\Lib\site-packages\torch\lib libiomp5md.dll 文件，报错解决了

主要Bug2

colmap报错

复制代码

F20240429 20:43:06.975742 42028 cache.h:132] Check failed: max_num_elems > 0 (0 vs. 0)
*** Check failure stack trace: ***
    @   00007FFBD6B16E5B  public: void __cdecl google::LogMessage::Flush(void) __ptr64
    @   00007FFBD6B15739  public: __cdecl google::LogMessageFatal::~LogMessageFatal(void) __ptr64
    @   00007FF6583B7B1A  (unknown)
    @   00007FF6583C8E0E  (unknown)
    @   00007FF6583BF99B  (unknown)
    @   00007FF6583AC130  (unknown)
    @   00007FF6583AA9CF  (unknown)
    @   00007FFC63E31BB2  _configthreadlocale
    @   00007FFC642E7344  BaseThreadInitThunk
    @   00007FFC661226B1  RtlUserThreadStart
Traceback (most recent call last):
  File "F:\ZXRworkSpace\PythonProject\BlindAss_Gen6D\Gen6D\prepare.py", line 98, in <module>
    build_colmap_model_no_pose(parse_database_name(args.database_name),args.colmap_path)
  File "F:\ZXRworkSpace\PythonProject\BlindAss_Gen6D\Gen6D\colmap_script.py", line 99, in build_colmap_model_no_pose
    extract_and_match_sift(colmap_path, database_path, image_path)
  File "F:\ZXRworkSpace\PythonProject\BlindAss_Gen6D\Gen6D\colmap_script.py", line 69, in extract_and_match_sift
    subprocess.run(cmd, check=True)
  File "C:\Users\StealMoon\.conda\envs\bass_gen\lib\subprocess.py", line 528, in run
    raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['F:\\ZXRworkSpace\\PythonProject\\BlindAss_Gen6D\\COLMAP3.8\\lib\\colmap.exe', 'exhaustive_matcher', '--database_path', 'data\\custom\\support\\colmap\\database.db']' returned non-zero exit status 3221226505.

当 COLMAP 无法找到姿势时，通常会出现此错误。它可能是由许多问题引起的，例如运动模糊过多、图像不足或场景中动态对象过多。

unity里面截的虚拟图像不能用 ,下载了作者的数据集，运行了一个小时左右

Writing output: data\custom\mouse\colmap\pointcloud.ply

执行成功的命令，针对博主的数据集能够成功执行

shell 复制代码

python prepare.py --action sfm --database_name custom/mouse --colmap F:\ZXRworkSpace\PythonProject\BlindAss_Gen6D\COLMAP3.8\COLMAP.bat

shell 复制代码

python prepare.py --action sfm --database_name custom/support --colmap F:\ZXRworkSpace\COLMAP3.8\lib\colmap.exe

问题依旧存在

This error typically arrises when COLMAP fails to find poses. It can be cause by a number of issues such as too much motion blur, not enough images, or too many dynamic objects in the scene.

看来问题确实是数据集，无法识别到物体的位姿。

python prepare.py --action sfm --database_name custom/support --colmap colmap.exe
需要注意的就是一定要准备一个合适的能偶识别出位姿的数据集

报错

加载模型、数据都没问题，读取数据库的时候出错了

复制代码

Traceback (most recent call last):
  File "F:\ZXRworkSpace\PythonProject\BlindAss_Gen6D\Gen6D\predict.py", line 97, in <module>
    main(args)
  File "F:\ZXRworkSpace\PythonProject\BlindAss_Gen6D\Gen6D\predict.py", line 32, in main
    estimator.build(ref_database, split_type='all')
  File "F:\ZXRworkSpace\PythonProject\BlindAss_Gen6D\Gen6D\estimator.py", line 140, in build
    object_center = get_object_center(database)
  File "F:\ZXRworkSpace\PythonProject\BlindAss_Gen6D\Gen6D\dataset\database.py", line 377, in get_object_center
    return database.center
AttributeError: 'CustomDatabase' object has no attribute 'center'

出错原因

这里应该要找数据库的center ，但是说找不到