shuxjweb / yet-another-reid-baseline

A simple baseline for Person ReID, it achieves 3rd place in VisDA2020 challenge.

The repo contains our code for VisDA 2020 challenge

What's new

• practical post-process: remove camera bias
• circle loss (class-level and pair-wise)
• memory bank
• Mix precision (FP16)
• Advanced Data augmentation: augmix, auto-augmentation
• pseudo label based method for unsupervised learning
• efficient search and re-rank by faiss
• Multi-GPU (single node DDP)
• SOTA benchmark
• Distillation

Requirement

1. pytorch>=1.2.0
2. yacs
3. sklearn
4. apex
5. faiss (pip install faiss-gpu)

Reproduce results on VisDA 2020 Challenge

Refer to VISDA20.md and tech_report, trained models can be download from here

• leaderboard (ranged by rank1)

team	mAP	rank1
vimar	76.56%	84.25%
xiangyu(ours)	72.39%	83.85%
yxge	74.78%	82.86%

• Ablation on validation set

method	mAP	rank1
personx-spgan	37.7%	63.7%
+pseudo label	51.8%	77.7%
+BN finetune	55.5%	81.4%
+re-rank	73.4%	80.9%
+remove camera bias	79.5%	89.1%
ensemble	82.7%	90.7%

Benchmark

Setting: ResNet50-ibn-a, single RTX 2080 Ti, FP16

• market1501

method	mAP	rank1
bag-of-tricks	88.2%	95.0%
fast reid	89.3%	95.3%
ours	88.4%	95.1%

• dukemtmc-reid

method	mAP	rank1
bag-of-tricks	79.1%	90.1%
fast-reid	81.2%	90.8%
ours	80.1%	90.3%

• msmt17(v2)

method	mAP	rank1
Bag of Tricks	54.4%	77.0%
fast reid	60.6%	83.9%
ours	60.6%	83.1%