xiezhy6 / GP-VTON

Official Implementation for CVPR2023 paper "GP-VTON: Towards General Purpose Virtual Try-on via Collaborative Local-Flow Global-Parsing Learning"
475 stars 83 forks source link

【Fatal】Why are some images missing in train_pairs_1008.txt and test_pairs_paired_1008.txt? #71

Open takesukeDS opened 8 months ago

takesukeDS commented 8 months ago

Thank you for your great work. The issue is described bellow.

The number of upper cloth images in your DressCode dataset is 15360.

(base) root@d2c7e4a7755e:/app/gp-vton/data/DressCode_512# ls upper/image | wc -l
15360

The sum of numbers of upper cloth images included in train_pairs_1008.txt and test_pairs_paired_1008.txt is not the same.

(base) root@d2c7e4a7755e:/app/gp-vton/data/DressCode_512# grep upper  test_pairs_paired_1008.txt | wc -l
1673
(base) root@d2c7e4a7755e:/app/gp-vton/data/DressCode_512# grep upper  train_pairs_1008.txt | wc -l
12505

Additionally, I believe the number of predefined test paires for unpaired setting is 1800, however you seem to use only 1559.

(base) root@d2c7e4a7755e:/app/gp-vton/data/DressCode_512# grep upper  test_pairs_unpaired_1008.txt | wc -l
1559

This should affected reported scores such as FID and LPIPS. Is there any valid reason you reduced the widely used dataset? Also the same issue for your vitonhd dataset.

Best regards