数据集路径配置的问题亦或是

[Lovebaby-1mt]

在运行 python pretrain_lm.py --cfg config/lm/init_gnn2_hrnetw18_npt512_lm.txt后出现的问题

Traceback (most recent call last): File "pretrain_lm.py", line 303, in main(configs) File "pretrain_lm.py", line 130, in main test_dataset = lm_dataset_single_obj_pytorch_code2d( File "/root/autodl-tmp/checkerpose/lm_dataset_pytorch.py", line 322, in init self.data_dicts = get_lm_data_dicts(dataset_name, dataset_root) File "/root/autodl-tmp/checkerpose/tools_for_LM/get_lm_datasets.py", line 402, in get_lm_data_dicts data_dicts = get_lm_13_dicts(data_cfg) File "/root/autodl-tmp/checkerpose/tools_for_LM/get_lm_datasets.py", line 95, in get_lm_13_dicts K = np.array(cam_dict[str_im_id]["cam_K"], dtype=np.float32).reshape(3, 3) KeyError: '0' 这个问题实在没有头绪，希望能得到解答！

[RuyiLian]

I do not know the exact problem based on the current information, but it seems that cam_dict does not have the key 0. I just verify that str_im_id=0 should be a valid key, then there may be something wrong about cam_dict. cam_dict is created as https://github.com/RuyiLian/CheckerPose/blob/fb725bed2a9eb6c3646c476c017d9ad00aa63a28/checkerpose/tools_for_LM/get_lm_datasets.py#L80 I suggest that you could check whether the path is invalid, or whether the json file is incomplete.

[Lovebaby-1mt]

Thank you for replying me After re-preparing the dataset I successfully ran the pretrain.I made a mistake in dataset,not download the BOPdataset but the officical,and mix up the lm and lmo.but now problem is fixed and sincere thanks for your warm reply and your sophisticated ieda.

[Lovebaby-1mt]

抱歉再次追问一下，我的lmo路径配置结构是 -lmo --fps_202212 --image_set --models --models_eval --test --train --camera.json --dataset_info.md --test_targets_all.json --test_targets_bop19.json

但最后发生了错误 datasets/BOP_DATASETS/lmo/models/obj_000001.ply datasets/BOP_DATASETS/lmo/models/obj_000005.ply datasets/BOP_DATASETS/lmo/models/obj_000006.ply datasets/BOP_DATASETS/lmo/models/obj_000008.ply datasets/BOP_DATASETS/lmo/models/obj_000009.ply datasets/BOP_DATASETS/lmo/models/obj_000010.ply datasets/BOP_DATASETS/lmo/models/obj_000011.ply datasets/BOP_DATASETS/lmo/models/obj_000012.ply if models are not fully listed above, please make sure there are ply files available datasets/BOP_DATASETS/lmo/models/obj_000001.ply obj_diameter 102.099 load FPS points from datasets/BOP_DATASETS/lmo/fps_202212/obj_000001.pkl FPS points, [before normalization] min -45.8286018371582 max 45.687198638916016 [after normalization] min -0.6510339126594673 max 0.9920106815910728 Traceback (most recent call last): File "pretrain.py", line 342, in main(configs) File "pretrain.py", line 114, in main print("training_data_folder image example:", rgb_files[obj_id][0], flush=True) IndexError: list index out of range

[RuyiLian]

We follow the conventions of previous work and thus it may be a bit confusing for LM and LMO dataset. When naming the training sets in our codes, for LM we follow GDR-Net while for LM-O we follow ZebraPose.

For LMO dataset, there are two sets of training data: train_pbr and train_real, as shown in the config file https://github.com/RuyiLian/CheckerPose/blob/fb725bed2a9eb6c3646c476c017d9ad00aa63a28/checkerpose/config/lmo/init_gnn2_hrnetw18_npt512.txt#L5-L6

• train_pbr can be downloaded from https://huggingface.co/datasets/bop-benchmark/datasets/resolve/main/lm/lm_train_pbr.zip (which contains physically based rendered images). Note that we use this synthetic training set in LMO but not in LM.
• train_real is actually the test images of LM. It can be downloaded from https://huggingface.co/datasets/bop-benchmark/datasets/resolve/main/lm/lm_test_all.zip. We use this set of images as real training images for LMO following previous works.

[Lovebaby-1mt]

你好，那关于ycbv的config是在哪呢，而且readme中提到的在ycbv数据集下的train_real是不是也来自于lm_test_all.zip

[RuyiLian]

1. For the configs of YCB-V dataset, I write a script convert_cfg_lmo_to_ycbv.py to generate configs from the corresponding ones of LM-O dataset (see README for more details of converting the configs). This is because for both LM-O and YBC-V, we train a single network for each object with almost the same hyper-parameters.
2. The data of YCB-V is completely independent of LM/LM-O. You can download train_pbr from https://huggingface.co/datasets/bop-benchmark/datasets/resolve/main/ycbv/ycbv_train_pbr.zip and train_real from https://bop.felk.cvut.cz/media/data/bop_datasets/ycbv_train_real.zip. To get more information of 6DoF object pose estimation datasets, you could check https://bop.felk.cvut.cz/datasets/.

[Lovebaby-1mt]

"And then run pretrain.py similar to LMO."的意思是运行 python pretrain.py --cfg config/ycbv/init_gnn2_hrnetw18_npt512.txt --obj_name ape

是这样的吗？但是我这边却报错 Traceback (most recent call last): File "pretrain.py", line 343, in main(configs) File "pretrain.py", line 82, in main obj_id = int(obj_name_obj_id[obj_name] - 1) # now the obj_id started from 0 KeyError: 'ape'

如果运行 python pretrain.py --cfg config/ycbv/init_gnn2_hrnetw18_npt512.txt

Traceback (most recent call last): File "pretrain.py", line 328, in check_point_path = check_point_path + config_file_name + '/' + args.obj_name TypeError: can only concatenate str (not "NoneType") to str

[RuyiLian]

Try to change --obj_name to one of the objects used in YCB-V dataset, e.g. master_chef_can, and run the command as: python pretrain.py --cfg config/ycbv/init_gnn2_hrnetw18_npt512.txt --obj_name master_chef_can The names of all the YCB-V objects can be found in https://github.com/RuyiLian/CheckerPose/blob/fb725bed2a9eb6c3646c476c017d9ad00aa63a28/checkerpose/tools_for_BOP/common_dataset_info.py#L30-L52

LM dataset uses a different set of objects as listed in
https://github.com/RuyiLian/CheckerPose/blob/fb725bed2a9eb6c3646c476c017d9ad00aa63a28/checkerpose/tools_for_LM/get_lm_datasets.py#L300-L314 and LM-O dataset uses 8 of the LM objects (i.e. ape, can, cat, driller, duck, eggbox, glue, holepuncher).

[Lovebaby-1mt]

how to test with trained models you upload

[RuyiLian]

please check https://github.com/RuyiLian/CheckerPose?tab=readme-ov-file#test-with-trained-model

[Lovebaby-1mt]

我遇到了一些问题： 用我自己训练的模型测试的话，路径会是checkerpose/experiments/checkpoints/hr18GNN2_res6_gnn3Skip_mlpQuery_lm/best_score/0_9638step239999

而下载的模型，比如说测试LM LM/archive/data.pkl 我这边运行不下去，觉得两边格式不完全一样，所以想问一下

[RuyiLian]

Could you try to use LM instead of LM/archive/data.pk for the checkpoint path? If it still does not work, make sure that the downloaded checkpoints are not decompressed.

[Lovebaby-1mt]

Sorry, is there any code for visualizing the results in this project?

[RuyiLian]

No, you could look into other GitHub repos of pose estimation.