.png)
[📢 Paper: ICMEW 2024] [📢 Poster]
dataset: UAVhuman-reid
The ultimate results of our method is 82.2(mAP), 81.3(Rank-1), 93.3(Rank-5).
First of all, create a conda env, then install packages detailed in enviroments.sh. Make sure that the cuda version of your machine matches the one in enviroments.sh (otherwise modify it into the one compatible with your machine).
conda create -n reid python==3.9
conda activate reid
bash enviroments.sh
Note that, all experiments are conducted using single GPU: NVIDIA Titan RTX or A6000.
model link: https://drive.google.com/drive/folders/15_afl7WAKBJMuGXQKsVM4VzWrwZcSs7R?usp=sharing
The first one is the pretrained model. Others are multiple models trained on uavhuman for ensemble.
We use the ViT-base/12 model pretrained on Market+DukeMTMC+MSMT17+cuhk02+cuhk03+CUHK-SYSU.
Furthermore, we add 7 tokens (like cls token) for attribute learning which largely improves the performance on uavhuman.
Modify the paths and settings in config/pretrained_vit.yml, then
python train.py --config_file config/#your_config_name#.yml
We use multiple tricks to re-rank the ranking lists of queries, like re-ranking, rank fusion of multiple models. Besides, we use the setting of multi-shot query images to further improve the performance.
Download models from Google Drive (or train them by yourself) and modify the model paths in test_ensemble.py as your trained model path, then
python test_ensemble.py
As detailed in test_ensemble.py, 5 distinct models are used for the ultimate performance. Their performances are shown below:
| multi-query setting | mAP | Rank-1 | Rank-5 | (mAP+R1)/2 |
|---|---|---|---|---|
| 1. ViT-B/12(imagenet) | 77.8 | 79.6 | 92.4 | 78.70 |
| 2. ViT-B/12(pretrained) | 79.7 | 80.4 | 93.3 | 80.05 |
| 3. 2+384x192 | 80.2 | 80.0 | 95.1 | 80.10 |
| 4. 3+soft REA | 79.0 | 79.1 | 92.6 | 79.05 |
| 5. 4+dirty samples filtering | 79.5 | 80.6 | 92.6 | 80.05 |
| 1+2+3+4+5 | 80.5 | 81.5 | 93.7 | 81.00 |
| 1+2+3+4+5(re-rank:4,4,0.45) | 82.2 | 81.3 | 93.3 | 81.75 |
As for the fifth one, we mannually select some dirty samples from the train-set (detailed in "old_new_name.json"), you can modify the path in the line 83 of data/datasets/uavhuman.py, then uncomment the codes related to dirty sample filtering and start training.
"1+2+3+4+5" indicates we ensemble the results got by 1, 2, 3, 4 and 5, follow test_ensemble.py for more information.