Lightweight-DNN-Triplet-Center-Loss-Person-Reid

Lightweight DNN using triplet center loss for Person re-identification benchmark on Market1501 dataset, the triplet center loss is based on this implementation.

Dependencies

• Tensorflow 2.3
• keras_self_attention
• efficientnet
• tensorflow-addons
• matplotlib
• numpy
• scikit-learn 0.23.1

Method

I build two types of model, the one that use one-stream network (without temporal block) and the other use two-stream network (requires 1 extension of input dimension), the overall architecture is based on this model but using single output instead of multi-output. The evaluation of dataset is followed the standard of Market1501 dataset.

Results

*approximate parameters based on their backbone

Benchmark References

[29]A. Hermans, L. Beyer, and B. Leibe, “In defense of the triplet loss for person re-identification,”
[23] Y. Sun, L. Zheng, W. Deng, and S. Wang, “SVDNet for Pedestrian Retrieval,”
[24] D. Li, X. Chen, Z. Zhang, and K. Huang, “Learning deep tr-Aware features over body and latent parts for person re-identification,” 
[25] T. Guo, D. Wang, Z. Jiang, A. Men, and Y. Zhou, “Deep Network with Spatial and Channel Attention for Person Re-identification,”
[27] X. Guo, “PERSON RE-IDENTIFICATION BY DEEP LEARNING MUTI-PART INFORMATION COMPLEMENTARY 
[28] M. Jiang, Z. Li, and J. Chen, “Person Re-Identification Using Color Features and CNN Features,”