search

adhirajghosh / RPTM_reid

[WACV'23] The official code for "Relation Preserving Triplet Mining for Stabilising the Triplet Loss in Re-identification Systems"
https://openaccess.thecvf.com/content/WACV2023/papers/Ghosh_Relation_Preserving_Triplet_Mining_for_Stabilising_the_Triplet_Loss_In_WACV_2023_paper.pdf
MIT License
29 stars 3 forks source link

mAP and your weight file ? #9

Open wjtan99 opened 3 weeks ago

wjtan99 commented 3 weeks ago

I trained as you suggested by changing TEST.EVAL to Flase, and adding TEST.TEST_SIZE=800.

python main.py --config_file configs/veri_r101.yml

I only got the follow results (not done yet, but the results are not expected to change much):

Train Epoch: [74][1801/1888] Time 0.199 (0.230) Loss 1.0569 (1.0329) Acc 100.00 (99.55) lr 5e-06
=> Test Evaluating veri ... Extracted features for query set, obtained 1678-by-2048 matrix Extracted features for gallery set, obtained 11579-by-2048 matrix => BatchTime(s)/BatchSize(img): 0.007/100 Computing CMC and mAP Results ---------- mAP: 65.7% CMC curve Rank-1 : 91.3% Rank-5 : 96.0% Rank-10 : 98.0%

Computing CMC and mAP Re-Ranked Results-- mAP: 76.4% CMC curve Rank-1 : 92.7% Rank-5 : 95.4% Rank-10 : 96.2%

This is way worse than your published results. Can you suggest what was wrong and can you share your weight file?

wjtan99 commented 3 weeks ago

I found that was due to the image size was set to 128*128. For this size, my results are better than the results in the Table 5 of the WACV paper.

adhirajghosh commented 3 weeks ago

Hi, thanks for your issue. May I close it now that you resolved it yourself?

wjtan99 commented 3 weeks ago

I still cannot get mAP=88 for 240×240. Can you share your weight file?

wjtan99 commented 2 weeks ago

I finished training input size = 240. However the mAP is about the same as for input size = 128. I am wondering how you got the results in Table 5. Can you share your weight files?
Thanks a lot.