The repo contains our code for VisDA 2020 challenge
- pytorch>=1.2.0
- yacs
- sklearn
- apex
- 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% |
Setting: ResNet50-ibn-a, single RTX 2080 Ti, FP16
method |
mAP |
rank1 |
bag-of-tricks |
88.2% |
95.0% |
fast reid |
89.3% |
95.3% |
ours |
88.4% |
95.1% |
method |
mAP |
rank1 |
bag-of-tricks |
79.1% |
90.1% |
fast-reid |
81.2% |
90.8% |
ours |
80.1% |
90.3% |
method |
mAP |
rank1 |
Bag of Tricks |
54.4% |
77.0% |
fast reid |
60.6% |
83.9% |
ours |
60.6% |
83.1% |