Implementation of paper 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.yaml

Test:

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.

Full Precision - Baseline mAP: 81.15 CMC1: 96.96 lambda:0.6 beta:1.0

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.

Half Precision - Baseline mAP: 91.44 CMC1: 86.72 lambda:0.2 beta:1.0

Hybrid R50+BoT

Please cite our paper if inspired on proposed techniques or code.

@INPROCEEDINGS{10422175,
  author={Almeida, Eurico and Silva, Bruno and Batista, Jorge},
  booktitle={2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC)}, 
  title={Strength in Diversity: Multi-Branch Representation Learning for Vehicle Re-Identification}, 
  year={2023},
  volume={},
  number={},
  pages={4690-4696},
  keywords={Representation learning;Metadata;Feature extraction;Real-time systems;Task analysis;Intelligent transportation systems;Residual neural networks},
  doi={10.1109/ITSC57777.2023.10422175}}

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.