Cosypose modification

Setup

仓库: https://github.com/Simple-Robotics/cosypose

git clone --recurse-submodules https://github.com/Simple-Robotics/cosypose.git
cd cosypose
conda env create -n cosypose --file environment.yaml

注意执行这一步的时候pip 会提示setuptools 和matplotlib-inline不符合3.7.6的python，到环境中手动安装适配的版本

conda activate cosypose
pip install setuptools==63.4.1
pip install matplotlib-inline==0.1.6

git lfs pull
python setup.py install
python setup.py develop

根据README下载数据
注意第一块指令无法下载成功，由 https://bop.felk.cvut.cz/datasets/ 得知下载链接迁移到了huggingface, https://huggingface.co/datasets/bop-benchmark/datasets/tree/main/ycbv 可以从这里手动下载测试集并放置到local_data/bop_datasets/ycbv/test

设置测试使用的models

cp ./local_data/bop_datasets/ycbv/model_bop_compat_eval ./local_data/bop_datasets/ycbv/models

Debug

np.where(mask)[0].item()

运行

export CUDA_VISIBLE_DEVICES=0 
python -m cosypose.scripts.run_cosypose_eval --config ycbv

时出现报错

Traceback (most recent call last):
  File "/home/cyl/.conda/envs/cosypose/lib/python3.7/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/home/cyl/.conda/envs/cosypose/lib/python3.7/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/home/cyl/cosypose/cosypose/scripts/run_cosypose_eval.py", line 491, in <module>
    main()
  File "/home/cyl/cosypose/cosypose/scripts/run_cosypose_eval.py", line 332, in main
    scene_ds = make_scene_dataset(ds_name)
  File "/home/cyl/cosypose/cosypose/datasets/datasets_cfg.py", line 68, in make_scene_dataset
    ids.append(np.where(mask)[0].item())
ValueError: can only convert an array of size 1 to a Python scalar

添加debug输出，得到

Debug - scene_id: 48, view_id: 1
Debug - mask matches: 1
Debug - where result shape: (1,), values: [225]
Debug - scene_id: 48, view_id: 36
Debug - mask matches: 1
Debug - where result shape: (1,), values: [226]
Debug - scene_id: 48, view_id: 47
Debug - mask matches: 1
Debug - where result shape: (1,), values: [227]
Debug - scene_id: 48, view_id: 83
Debug - mask matches: 1
Debug - where result shape: (1,), values: [228]
Debug - scene_id: 48, view_id: 112
Debug - mask matches: 1
Debug - where result shape: (1,), values: [229]
Debug - scene_id: 48, view_id: 135
Debug - mask matches: 0
Debug - where result shape: (0,), values: []
0:00:00.912023 - Expected exactly one match, got 0 matches for scene_id=48, view_id=135

发现是下载的测试数据集并不包含数据集keyframe.txt中所有的帧，导致一些关键帧识别不到

运行到一半被终止的情况

如果想重新开始新的训练： 清空local_data/joblib_cache

Framework

prediction script cosypose.scripts.run_cosypose_eval

AI explanation

The script predicts object poses based on multi-view input by following these steps:

1. Dataset Loading: It first loads the dataset using the make_scene_dataset function, which prepares the scene data for evaluation. The dataset is wrapped in a MultiViewWrapper to handle multiple views.

2. Model Loading: The script loads pre-trained models for pose prediction using the load_models function. It loads both coarse and refiner models based on the configuration specified in the command-line arguments.

3. Prediction Setup: The script sets up the prediction parameters, including the number of iterations for coarse and refiner models, and whether to skip multi-view processing based on the number of views specified.

4. Multi-view Prediction: The MultiviewScenePredictor is initialized with the mesh database, which is used to predict poses across multiple views. The MultiviewPredictionRunner is then used to run predictions on the dataset, leveraging the multi-view setup to improve pose estimation accuracy.

5. Pose Estimation: The script uses the loaded models to predict object poses. It processes detections from either pix2pose or posecnn depending on the dataset, and refines these predictions using the refiner model.

6. Evaluation: After predictions, the script evaluates the predicted poses using the PoseEvaluation class. It calculates various metrics like ADD-S and AUC to assess the accuracy of the pose predictions.

7. Results Logging: Finally, the script logs the results, including evaluation metrics, and saves them to a specified directory.

The multi-view approach allows the script to leverage information from different viewpoints, which can help resolve ambiguities and improve the robustness of the pose estimation.

Model dataset for the initialization of MultiviewScenePredictor

In run_cosypose_eval we initialize MultiviewScenePredictor in this way:

mv_predictor = MultiviewScenePredictor(mesh_db)

In the MultiviewScenePredictor we use the mesh_db to initialize MultiviewRefinement and solve:

problem = MultiviewRefinement(candidates=candidates_n,
                    cameras=cameras,
	                pairs_TC1C2=pairs_TC1C2,
	                mesh_db=self.mesh_db_ba)
ba_outputs = problem.solve(
	n_iterations=ba_n_iter,
	optimize_cameras=not use_known_camera_poses,
)

The solve function of MultiviewRefinement:

def solve(self, sample_n_init=1, **lm_kwargs):
	timer_init = Timer()
	timer_opt = Timer()
	timer_misc = Timer()

	timer_init.start()
	TWO_9d_init, TCW_9d_init = self.robust_initialization_TWO_TCW(n_init=sample_n_init)
	timer_init.pause()

	timer_opt.start()
	TWO_9d_opt, TCW_9d_opt, history = self.optimize_lm(
		TWO_9d_init, TCW_9d_init, **lm_kwargs)
	timer_opt.pause()

	timer_misc.start()
	objects, cameras = self.make_scene_infos(TWO_9d_opt, TCW_9d_opt)
	objects_init, cameras_init = self.make_scene_infos(TWO_9d_init, TCW_9d_init)
	history = self.convert_history(history)
	timer_misc.pause()

	outputs = dict(
		objects_init=objects_init,
		cameras_init=cameras_init,
		objects=objects,
		cameras=cameras,
		history=history,
		time_init=timer_init.stop(),
		time_opt=timer_opt.stop(),
		time_misc=timer_misc.stop(),
	)
	return outputs

Adaption

不急