search

mangye16 / DDAG

Pytorch Code of DDAG for Visible-Infrared Person Re-Identification (ECCV20)
MIT License
76 stars 16 forks source link

The performance of DDAG on RegDB #5

Open yjh576 opened 3 years ago

yjh576 commented 3 years ago

This paper reports that DDGA is evaluated under both visible-to-infrared and infrared-to-visible query settings. I re-implemented the code you published without making any changes. I obtain 31.25% map under visible to infrared setting, which is much more lower than 61.80% map reported in your paper. Under infrared to visible setting, similar results (61.97 vs 63.46) is still found.

Thanks in advances.

mangye16 commented 3 years ago

I suppose you might have made a mistake in the visible-to-infrared setting.

Just keep all the settings as default (Do not change the test_mode in [Line 69]), since this represents the feature extraction mode.

The only thing you need to change is the distance computation, changing the query and gallery setting.

yjh576 commented 3 years ago

The visible-to-infrared setting is that query_img, query_label = process_test_regdb(data_path, trial=args.trial, modal='visible') gall_img, gall_label = process_test_regdb(data_path, trial=args.trial, modal='thermal'). The infrared-to-visible setting is that
query_img, query_label = process_test_regdb(data_path, trial=args.trial, modal='thermal'') gall_img, gall_label = process_test_regdb(data_path, trial=args.trial, modal='visible'). Do you mean it?

Testing trial is only for RegDB dataset. Can you explain it in detail？

mangye16 commented 3 years ago

You may refer to the baseline code at test.py in Line 254.

Other parts are kept the same.