haochenheheda
Can not revise same performance on DukeMTMC
I use the config file same to Market1501 provided. But I get 78.4(mAP) and 86.2 R-1. In your paper, the mAP and R-1 with nformer is 83.5 and 89.4. So is the config file for dukemtmc is different with market1051?
run command: python3 tools/train.py --config_file='configs/softmax_triplet_with_center.yml' MODEL.DEVICE_ID "('0')" DATASETS.NAMES "('dukemtmc')" DATASETS.ROOT_DIR "('~/workplace/dataset')" OUTPUT_DIR "('./log/dukemtmc-nformer')"
log: 2022-10-21 23:22:36,387 reid_baseline.train INFO: Epoch[120] Iteration[20/224] Loss: 1.055, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:22:39,391 reid_baseline.train INFO: Epoch[120] Iteration[40/224] Loss: 1.054, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:22:42,403 reid_baseline.train INFO: Epoch[120] Iteration[60/224] Loss: 1.052, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:22:45,415 reid_baseline.train INFO: Epoch[120] Iteration[80/224] Loss: 1.051, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:22:48,429 reid_baseline.train INFO: Epoch[120] Iteration[100/224] Loss: 1.051, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:22:51,458 reid_baseline.train INFO: Epoch[120] Iteration[120/224] Loss: 1.051, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:22:54,501 reid_baseline.train INFO: Epoch[120] Iteration[140/224] Loss: 1.049, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:22:57,544 reid_baseline.train INFO: Epoch[120] Iteration[160/224] Loss: 1.048, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:23:00,590 reid_baseline.train INFO: Epoch[120] Iteration[180/224] Loss: 1.047, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:23:03,604 reid_baseline.train INFO: Epoch[120] Iteration[200/224] Loss: 1.046, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:23:06,572 reid_baseline.train INFO: Epoch[120] Iteration[220/224] Loss: 1.043, Acc: 1.000, Base Lr: 3.50e-06 2022-10-21 23:23:07,325 reid_baseline.train INFO: Epoch 120 done. Time per batch: 30.887[s] Speed: 464.2[samples/s] 2022-10-21 23:23:07,326 reid_baseline.train INFO: ---------- 2022-10-21 23:24:19,715 reid_baseline.train INFO: Validation Results - Epoch: 120 2022-10-21 23:24:19,716 reid_baseline.train INFO: mAP: 69.0% 2022-10-21 23:24:19,716 reid_baseline.train INFO: CMC curve, Rank-1 :83.7% 2022-10-21 23:24:19,716 reid_baseline.train INFO: CMC curve, Rank-5 :92.3% 2022-10-21 23:24:19,716 reid_baseline.train INFO: CMC curve, Rank-10 :93.9% 2022-10-21 23:32:57,497 reid_baseline.train INFO: mAP: 77.1% 2022-10-21 23:32:57,497 reid_baseline.train INFO: CMC curve, Rank-1 :86.0% 2022-10-21 23:32:57,497 reid_baseline.train INFO: CMC curve, Rank-5 :92.2% 2022-10-21 23:32:57,497 reid_baseline.train INFO: CMC curve, Rank-10 :93.9% 2022-10-21 23:41:21,619 reid_baseline.train INFO: mAP: 78.3% 2022-10-21 23:41:21,620 reid_baseline.train INFO: CMC curve, Rank-1 :85.6% 2022-10-21 23:41:21,620 reid_baseline.train INFO: CMC curve, Rank-5 :92.5% 2022-10-21 23:41:21,620 reid_baseline.train INFO: CMC curve, Rank-10 :94.1% 2022-10-21 23:49:46,520 reid_baseline.train INFO: mAP: 78.4% 2022-10-21 23:49:46,520 reid_baseline.train INFO: CMC curve, Rank-1 :85.6% 2022-10-21 23:49:46,521 reid_baseline.train INFO: CMC curve, Rank-5 :92.1% 2022-10-21 23:49:46,521 reid_baseline.train INFO: CMC curve, Rank-10 :94.0% 2022-10-21 23:58:12,988 reid_baseline.train INFO: mAP: 78.4% 2022-10-21 23:58:12,989 reid_baseline.train INFO: CMC curve, Rank-1 :86.2% 2022-10-21 23:58:12,989 reid_baseline.train INFO: CMC curve, Rank-5 :92.3% 2022-10-21 23:58:12,989 reid_baseline.train INFO: CMC curve, Rank-10 :94.3%
I think maybe the feature extractor is not well-trained. Please check the performance with only the feature extractor. The mAP should be about 76 without NFormer. The training of the feature extractor is copied from this repo.
Do you mean that I can use Res50 weights pretrained by https://github.com/michuanhaohao/reid-strong-baseline. And could you provide your NFormer weights?
Dear author, I have been studying related work recently. Could you please provide me with your training model weights? Thank you very much!
