search

hijune6 / Hetero-center-triplet-loss-for-VT-Re-ID

MIT License
45 stars 5 forks source link

RegDB dataset: Lower value on mAP of the thermal to visible test mode #2

Open Ning5195 opened 3 years ago

Ning5195 commented 3 years ago

Nice job! Thanks for sharing the code. I run the train_mine.py on the RegDB dataset and got the model named 'regdb_c_tri_pcb_on_w_tri_2.0_s6_f256_share_net2_base_gm10_k4_p8_lr_0.1_seed_0_trial_1_best.t'. When test in the mode of visible to thermal, I got the results as follows: All Average: POOL: Rank-1: 94.87% | Rank-5: 96.97% | Rank-10: 98.21%| Rank-20: 99.20%| mAP: 90.33%| mINP: 83.49% However, when test in the mode of thermal to visible, I got lower value than that described in the paper as follows: All Average: POOL: Rank-1: 52.56% | Rank-5: 63.18% | Rank-10: 69.19%| Rank-20: 75.85%| mAP: 47.53%| mINP: 33.94% I switched the mode of test by changed the code in the line 84 of test_mine_pcb.py, as 'test_mode = [2, 1]' for visible to thermal and 'test_mode = [1, 2]' for thermal to visible. May I ask where I am making a mistake?

hijune6 commented 3 years ago

thanks for your interesting of our works. there is a parameter '--tvsearch' , may help u.发自我的华为手机-------- 原始邮件 --------发件人： Ning5195 notifications@github.com日期： 2020年12月31日周四 17:08收件人： hijune6/Hetero-center-triplet-loss-for-VT-Re-ID Hetero-center-triplet-loss-for-VT-Re-ID@noreply.github.com抄送： Subscribed subscribed@noreply.github.com主 题： [hijune6/Hetero-center-triplet-loss-for-VT-Re-ID] RegDB dataset: Lower value on mAP of the thermal to visible test mode (#2) Nice job! Thanks for sharing the code. I run the train_mine.py on the RegDB dataset and got the model named 'regdb_c_tri_pcb_on_w_tri_2.0_s6_f256_share_net2_base_gm10_k4_p8_lr_0.1_seed_0_trial_1_best.t'. When test in the mode of visible to thermal, I got the results as follows: All Average: POOL: Rank-1: 94.87% | Rank-5: 96.97% | Rank-10: 98.21%| Rank-20: 99.20%| mAP: 90.33%| mINP: 83.49% However, when test in the mode of thermal to visible, I got lower value than that described in the paper as follows: All Average: POOL: Rank-1: 52.56% | Rank-5: 63.18% | Rank-10: 69.19%| Rank-20: 75.85%| mAP: 47.53%| mINP: 33.94% I switched the mode of test by changed the code in the line 84 of test_mine_pcb.py, as 'test_mode = [2, 1]' for visible to thermal and 'test_mode = [1, 2]' for thermal to visible. May I ask where I am making a mistake?

—You are receiving this because you are subscribed to this thread.Reply to this email directly, view it on GitHub, or unsubscribe.

Ning5195 commented 3 years ago

Thank you for reply. It did work. Sorry I didn’t check the description of parameters carefully before.

And I have met another confusion. As I set P = 6, K = 8 and λ = 1.0 for the SYSU-MM01 dataset, and the model named 'sysu_c_tri_pcb_on_w_tri_1.0_s6_f256_share_net2_base_gm10_k8_p6_lr_0.1_seed_0_best.t' has got after running the code of train_mine.py. But I only got the results as follows. For all search, the results are: All Average: POOL: Rank-1: 40.17% | Rank-5: 69.29% | Rank-10: 80.65%| Rank-20: 89.92%| mAP: 38.31%| mINP: 23.15% For indoor search, the results are: All Average: POOL: Rank-1: 43.14% | Rank-5: 73.88% | Rank-10: 84.71%| Rank-20: 92.85%| mAP: 51.43%| mINP: 46.15%

So, should I set a larger batch size? Or there are other parameters I set incorrectly?