Implementation of the paper Unsupervised Domain Adaptive Re-Identification: Theory and Practice.
The selftraining scheme proposed in the paper is simple yet effective.
- Datasets (source dataset and target dataset).
- Pre-trained (on source dataset) model.
- PyTorch
To replicate the results in the paper, you can download pre-trained models on Market1501 from GoogleDrive and on DukeMTMC from GoogleDrive. Our models are trained with PyTorch 0.3.
python selftraining.py \
--src_dataset <name_of_source_dataset>\
--tgt_dataset <name_of_target_dataset>\
--resume <dir_of_source_trained_model>\
--data_dir <dir_of_source_target_data>\
--logs_dir <dir_to_save_model_after_adaptation>
dw_example.ipynb
is the file for replicating Figure 6 in the paper.
Rank-1 | Rank-5 | Rank-10 | mAP | |
---|---|---|---|---|
On source (DukeMTMC) | 80.8 | 91.2 | 94.2 | 65.4 |
On target (Market1501) | 46.8 | 64.6 | 71.5 | 19.1 |
After adaptation | 75.8 | 89.5 | 93.2 | 53.7 |
Rank-1 | Rank-5 | Rank-10 | mAP | |
---|---|---|---|---|
On source (Market1501) | 91.6 | 97.1 | 98.5 | 78.2 |
On target (DukeMTMC) | 27.3 | 41.2 | 47.1 | 11.9 |
After adaptation | 68.4 | 80.1 | 83.5 | 49.0 |
Our code is based on open-reid.