topo-epfl / crossloc Goto Github PK

Hi scholars, hello everyone. When I am executing -- task semantics/depth/normal, regardless of whether I change the weight name you provided to mlr_encoder_ * or ckpt_iter_ *, the following error still appears, indicating that the information was not read correctly. How should this situation be resolved? If I execute -- task coord, this problem won't occur. Why is this?
"(crossloc) user@user:~/workspaces/CrossLoc$ python3 test_single_task.py naturescape --task semantics --uncertainty mle --section test_drone_real --network_in /media/user/Data/CrossLoc-Datasets/ckpt_CrossLoc_network_weights/naturescape/in-place/
Namespace(fullsize=False, grayscale=False, hypotheses=64, inlieralpha=100, keywords=None, max_ckpt_iter=None, maxpixelerror=100, min_ckpt_iter=None, network_in='/media/slby/user/CrossLoc-Datasets/ckpt_CrossLoc_network_weights/naturescape/in-place/', plot=False, save_pred=False, scene='naturescape', search_dir=False, section=['test_drone_real'], session='', task='semantics', threshold=10, tiny=False, uncertainty='MLE')
To load network weight from /media/user/Data/CrossLoc-Datasets/ckpt_CrossLoc_network_weights/naturescape/in-place
With the keywords None, and min ckpt iter None, max ckpt iter None the following 1 network weight paths are retrieved:
Network weight #0: /media/user/Data/CrossLoc-Datasets/ckpt_CrossLoc_network_weights/naturescape/in-place/ckpt_iter_crossloc_se_naturescape_in-place.net
==================== Evaluating network #1 / 1 ====================
Path: /media/user/Data/CrossLoc-Datasets/ckpt_CrossLoc_network_weights/naturescape/in-place/ckpt_iter_crossloc_se_naturescape_in-place.net
Traceback (most recent call last):
File "test_single_task.py", line 420, in
main()
File "test_single_task.py", line 299, in main
network = config_network(scene, task, tiny, grayscale, uncertainty, fullsize, network_path, num_enc=num_enc)
File "/home/user/workspaces/CrossLoc/utils/evaluation.py", line 162, in config_network
raise NotImplementedError
NotImplementedError"

Hi,

Thanks for the code, I'm very interested in your outstanding work!!!
I have tried to retrain your network, and the result shows that the index values of in-place localization accuracy I reproduced over urbanscape dataset are better than those given in your paper. So I want to ask you what might cause this situation.

Looking forward to your reply.

Here is the result over urbanscape dataset from your paper:

Here is quantitative comparison between paper's result and my reproduced result: