The PyTorch implementation for jointly training of pedestrian attribute recognition and Re-ID. The model is evaluated on Market-1501(with attribute), DukeMTMC-reID(with attribute) datasets and achieve the state-of-the-art performance on the two tasks.
Jointly training with some hyper-parameters.
python3 train_joint.py [--warm_epoch 10] [--stride 1] [--erasing_p 0.5] [--batch-size 8] [--lr 0.02] [--num-epoch 80]
Evaluation for the pedestrian attribute recognition task.
python3 test_joint.py [--stride 1] [--fusion] [--alpha 0.5]
Evaluation for the Re-ID task.
python3 test_reid.py --stride 1 --frame_junk
Visualize the similarity of positive and negative samples.
python3 prepare_threshold.py
python3 threshold.py [--stride 1]
Recognize the pedestrian attributes in the given images.
python3 inference.py path_to_image [--stride 1]
Evaluation for reID on Market-1501
(Camera junk) Rank@1:0.930523 Rank@5:0.975653 Rank@10:0.983373 mAP:0.816014
(Frame junk) Rank@1:0.964371 Rank@5:0.987233 Rank@10:0.992280 mAP:0.843044
Evaluation for reID on DukeMTMC-reID
(Camera junk) Rank@1:0.853680 Rank@5:0.925494 Rank@10:0.948384 mAP:0.707455
(Frame junk) Rank@1:0.965889 Rank@5:0.987881 Rank@10:0.991023 mAP:0.798274
Evaluation for pedestrian attribute recognition on Market-1501-attribute
+------------+----------+-----------+--------+----------+
| attribute | accuracy | precision | recall | f1 score |
+------------+----------+-----------+--------+----------+
| young | 0.998 | 0.000 | 0.000 | 0.000 |
| teenager | 0.873 | 0.938 | 0.915 | 0.926 |
| adult | 0.886 | 0.526 | 0.553 | 0.539 |
| old | 0.994 | 0.000 | 0.000 | 0.000 |
| backpack | 0.877 | 0.818 | 0.656 | 0.728 |
| bag | 0.785 | 0.568 | 0.478 | 0.519 |
| handbag | 0.898 | 0.320 | 0.072 | 0.118 |
| clothes | 0.926 | 0.936 | 0.983 | 0.959 |
| down | 0.941 | 0.976 | 0.934 | 0.955 |
| up | 0.935 | 0.935 | 1.000 | 0.967 |
| hair | 0.893 | 0.867 | 0.832 | 0.849 |
| hat | 0.977 | 0.783 | 0.287 | 0.420 |
| gender | 0.931 | 0.945 | 0.895 | 0.919 |
| upblack | 0.953 | 0.900 | 0.734 | 0.809 |
| upwhite | 0.920 | 0.861 | 0.834 | 0.847 |
| upred | 0.972 | 0.889 | 0.832 | 0.859 |
| uppurple | 0.986 | 0.716 | 0.807 | 0.759 |
| upyellow | 0.976 | 0.921 | 0.801 | 0.857 |
| upgray | 0.911 | 0.835 | 0.419 | 0.558 |
| upblue | 0.948 | 0.882 | 0.443 | 0.590 |
| upgreen | 0.969 | 0.810 | 0.739 | 0.773 |
| downblack | 0.889 | 0.854 | 0.859 | 0.857 |
| downwhite | 0.960 | 0.677 | 0.508 | 0.580 |
| downpink | 0.988 | 0.846 | 0.664 | 0.744 |
| downpurple | 1.000 | 0.000 | 0.000 | 0.000 |
| downyellow | 0.999 | 0.000 | 0.000 | 0.000 |
| downgray | 0.873 | 0.765 | 0.393 | 0.519 |
| downblue | 0.862 | 0.782 | 0.430 | 0.555 |
| downgreen | 0.973 | 1.000 | 0.022 | 0.044 |
| downbrown | 0.962 | 0.779 | 0.625 | 0.694 |
+------------+----------+-----------+--------+----------+
Average accuracy: 0.9351
Average f1 score: 0.5982
Evaluation for pedestrian attribute recognition on DukeMTMC-reID-attribute
+-----------+----------+-----------+--------+----------+
| attribute | accuracy | precision | recall | f1 score |
+-----------+----------+-----------+--------+----------+
| backpack | 0.813 | 0.808 | 0.862 | 0.835 |
| bag | 0.807 | 0.402 | 0.366 | 0.383 |
| handbag | 0.934 | 0.170 | 0.008 | 0.015 |
| boots | 0.908 | 0.808 | 0.771 | 0.789 |
| gender | 0.873 | 0.842 | 0.823 | 0.832 |
| hat | 0.910 | 0.862 | 0.756 | 0.806 |
| shoes | 0.922 | 0.760 | 0.480 | 0.588 |
| top | 0.866 | 0.447 | 0.412 | 0.429 |
| upblack | 0.825 | 0.853 | 0.870 | 0.862 |
| upwhite | 0.957 | 0.775 | 0.428 | 0.552 |
| upred | 0.974 | 0.784 | 0.602 | 0.681 |
| uppurple | 0.996 | 0.077 | 0.015 | 0.026 |
| upgray | 0.903 | 0.643 | 0.332 | 0.437 |
| upblue | 0.943 | 0.779 | 0.502 | 0.611 |
| upgreen | 0.975 | 0.462 | 0.350 | 0.398 |
| upbrown | 0.980 | 0.459 | 0.212 | 0.290 |
| downblack | 0.772 | 0.741 | 0.753 | 0.747 |
| downwhite | 0.942 | 0.790 | 0.317 | 0.453 |
| downred | 0.988 | 0.684 | 0.436 | 0.532 |
| downgray | 0.917 | 0.355 | 0.212 | 0.265 |
| downblue | 0.778 | 0.709 | 0.598 | 0.649 |
| downgreen | 0.997 | 0.000 | 0.000 | 0.000 |
| downbrown | 0.976 | 0.892 | 0.499 | 0.640 |
+-----------+----------+-----------+--------+----------+
Average accuracy: 0.9111
Average f1 score: 0.5139
- 2019-12-4: upload the project
- 2019-12-9: add new transforms and fea_cat
- 2019-12-14: 1. add tricks: stride, erasing_p, warm_epoch, lr 2. add training option: continue or restart 3. fix the bug of image size among training and testing files 4. transform the dimension of features from 2048 to 512: remove classifier_re id.classifier but not classifier_reid
- 2019-12-16: add --frame_junk
- 2020-1-12: add --fusion --alpha
- 2020-1-13: add threshold evaluation
