how to generate density_maps_constant15 for vis
Closed this issue · 3 comments
zhou-rui1 commented
Hi, I am wondering about how to create this file in the vis_densityMap process:
Line 77 in 58c8d0c
gt_dmap_path = image_path.replace('.jpg', '.npy').replace('images', 'density_maps_constant15')
wfs123456 commented
Hello! DM-Count-loss only needs dot-map as ground truth. In vis_density-map.py，the ---constant15 is the density-map generated by Gaussian smoothing，referencing paper MCNN/CSRNet，and the constant15 is only used to compare the output of CCTrans. So, you can also delete this line of code，only print and visualize the output of CCTrans.
…---Original---
From: ***@***.***>
Date: Mon, Apr 4, 2022 23:09 PM
To: ***@***.***>;
Cc: ***@***.***>;
Subject: [wfs123456/CCTrans] how to generate density_maps_constant15 for vis (Issue #6)
Hi, I am wondering about how to create this file in the vis_densityMap process:
https://github.com/wfs123456/CCTrans/blob/58c8d0c257774bfd37b64cafbf265e5def344466/vis_densityMap.py#L77
gt_dmap_path = image_path.replace('.jpg', '.npy').replace('images', 'density_maps_constant15')
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>
zhou-rui1 commented
Thanks a lot! It goes well now.
wfs123456 commented
For NWPU dataset，i reshaped the resolution of all images to avoid out of memory of gpu. The preprocess_dataset_nwpu.py has been changed, that might have been handled incorrectly because of time. (please refer to the DM-Count code or directly adopt original resolution for training in NWPU if the gpu is enough!
…---Original---
From: ***@***.***>
Date: Sat, Apr 9, 2022 00:05 AM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [wfs123456/CCTrans] how to generate density_maps_constant15 forvis (Issue #6)
Hi, have you try the code on NWPU dataset? the validated mae seems to remain high(over 300) and it is hard to train lower, I am wondering if this is due to some problem in the preprocess_dataset_nwpu.py?
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: ***@***.***>