vehicle_reid_itsc2023
Strength in Diversity: Multi-Branch Representation Learning for Vehicle Re-Identification
Version 1.0
All models are trained with CUDA 11.3 and PyTorch 1.11 on RTX4090 and Ryzen 7950X.
Resuls are displayed as mAP / CMC1 in percentage values %.
VehicleID was not available at the time of writing, we report values for VehicleID now. We follow evaluation as FastReid with 10-fold cross-validation to select queries and gallery.
To train:
python main.py --model_arch MBR_4G --config ./config/config_BoT_VERIWILD.yamlTest:
python teste.py --path_weights ./logs/VERIWILD/MBR_4G/0/If you want to test I share some of the weights.
After code reformultation there are some slight changes to the values from what is written on the paper.
Veri-776
Full Precision - Baseline mAP: 81.15 CMC1: 96.96 lambda:0.6 beta:1.0
| R50 | 4G | 2G | 2X2G | 2X | 4X |
|---|---|---|---|---|---|
| CE + Tri | 82.81/97.38 | 83.04/97.14 | 83.67/97.32 | 81.82/96.96 | 82.31/97.32 |
| CE / Tri | 82.47/96.84 | 83.26/97.02 | 84.22/97.02 | 83.67/97.50 | 83.89/97.5 |
| R50+BoT | 4G | 2G | 2X2G | 2X | 4X |
|---|---|---|---|---|---|
| CE + Tri | 82.04/96.96 | 81.14/ 97.02 | 82.02/96.78 | 82.82/97.20 | 83.3/97.62 |
| CE / Tri | 82.67/97.02 | NULL | 82.57/97.32 | NULL | 84.72/97.68 |
Veri-Wild
Half Precision - Baseline mAP: 87.24 CMC1: 96.65 lambda:0.8 beta:1.0 Some values were updated after detecting weird behaviours with nn.parallel usage.
| R50 | 4G | 2G | 2X2G | 2X | 4X |
|---|---|---|---|---|---|
| CE + Tri | 85.16/94.81 | 86.66/96.08 | 86.7/95.52 | 87.64/96.39 | 87.31/96.05 |
| CE / Tri | 84.05/93.41 | 86.11/95.28 | 86.91/95.78 | 87.31/95.98 | 87.73/96.02 |
| R50+BoT | 4G | 2G | 2X2G | 2X | 4X |
|---|---|---|---|---|---|
| CE + Tri | 86.07/95.58 | 86.92/96.29 | 87.11/95.62 | 88.57/96.79 | 88.9/96.55 |
| CE / Tri | 85.29/94.85 | NULL | 86.52/95.21 | NULL | 86.9/95.75 |
VehicleID
Half Precision - Baseline mAP: 91.44 CMC1: 86.72 lambda:0.2 beta:1.0
| R50 | 4G | 2G | 2X2G | 2X | 4X |
|---|---|---|---|---|---|
| CE + Tri | 91.5/86.81 | 91.86/87.35 | 91.6/86.95 | 92.04/87.62 | 91.79/87.28 |
| CE / Tri | 90.68/85.56 | 90.94/85.59 | 91.44/86.66 | 91.45/86.83 | 91.91/87.36 |
Hybrid R50+BoT
| R50+BoT | 4G | 2G | 2X2G | 2X | 4X |
|---|---|---|---|---|---|
| CE + Tri | 91.35/86.46 | 91.66/87.01 | 91.48/86.76 | 91.99/87.39 | 92.03/87.49 |
| CE / Tri | 90.36/85.17 | NULL | 91.15/86.22 | NULL | 92.75/88.46 |
Please cite our paper if inspired on proposed techniques or code. BibTex reference will be here once the paper has been officially published in IEEE, meanwhile if you want you can use arxiv.preprint
Some of the code is reused from:
Parsing-based View-aware Embedding Network for Vehicle Re-Identification
Bag of Tricks and A Strong Baseline for Deep Person Re-identification
FastREID
So please also cite and support their work.