search

felja633 / RLLReg

Registration Loss Learning for Deep Probabilistic Point Set Registration
GNU General Public License v3.0
40 stars 4 forks source link

Some question about evaluation_threedmatch.py #1

Open Yellowshuohahaha opened 3 years ago

Yellowshuohahaha commented 3 years ago

I download RGB-D scenes from 3dmatch and pre-trained models. After set pretrained_networks and 3dmatch dataset to the path of that folder in config.py, I run the following script: python evaluation_threedmatch.py => loading checkpoint '/home/s206/disk/pycharmproject/RLLReg-master/data/models/ResUNetBN2C-feat32-3dmatch-v0.05.pth' => Setting voxel size to 0.05 => loading finished generate samples... evaluate batch 0 ... RLLReg resample time: 449.2 ms extract time: 644.6 ms Traceback (most recent call last): File "evaluation_threedmatch.py", line 140, in max_err_R=4.0, max_err_t=0.3, success_R=4, success_t=0.1, save_errors=True, plot=False) File "/home/s206/disk/pycharmproject/RLLReg-master/evaluation/evaluate.py", line 36, in benchmark errors[m.params.name].append(evaluate_registration(m, batch, vis, save=save)) File "/home/s206/disk/pycharmproject/RLLReg-master/evaluation/evaluate.py", line 10, in evaluate_registration out = method(data) File "/home/s206/anaconda3/envs/rllreg/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, **kwargs) File "/home/s206/disk/pycharmproject/RLLReg-master/models/rll_model_fcgf.py", line 164, in forward assert f.shape[0] == inds_list[ind_cnt].shape[0] AssertionError It seems the code doesn't run rightly. My torch version is 1.7 and minkowskiEngine version is 0.4. Can you help me?

felja633 commented 3 years ago

I see, it is probably some changes in Minkowski engine that makes my indexing to break. I will try to make this part more robust when I have the time. Thank you for pointing it out

Yellowshuohahaha commented 3 years ago

I see, it is probably some changes in Minkowski engine that makes my indexing to break. I will try to make this part more robust when I have the time. Thank you for pointing it out

Thank you very much! Looking forword to your work.

Yellowshuohahaha commented 2 years ago

Hi @felja633 , may I ask if reducing the Minkowski engine version can solve this problem? I want to use this project to do some engineering point cloud registration applications. Thanks.

felja633 commented 2 years ago

That could solve the problem. Try checking out a version from last year, at least that has worked for me in the passed