fast-reid/projects/StrongerBaseline/README.md

Stronger Baseline in FastReID

Training

To train a model, run

CUDA_VISIBLE_DEVICES=gpus python train_net.py --config-file <config.yaml>

For example, to launch a end-to-end baseline training on market1501 dataset with ibn-net on 4 GPUs, one should excute:

CUDA_VISIBLE_DEVICES=0,1,2,3 python train_net.py --config-file='configs/sbs_market1501.yml'

Experimental Results

stronger baseline tricks:

1. Non-local block
2. GeM pooling
3. Circle loss
4. Freeze backbone training
5. Cutout data augmentation & Auto Augmentation
6. Cosine annealing learning rate decay
7. Soft margin triplet loss

Market1501 dataset

Method	Pretrained	Rank@1	mAP	mINP
stronger baseline(ResNet50-ibn)	ImageNet	95.5	88.4	65.8
Robust-ReID	ImageNet	96.2	89.7	-

DukeMTMC dataset

Method	Pretrained	Rank@1	mAP	mINP
stronger baseline(ResNet50-ibn)	ImageNet	91.3	81.6	47.6
Robust-ReID	ImageNet	89.8	80.3	-

MSMT17 dataset

Method	Pretrained	Rank@1	mAP	mINP
stronger baseline(ResNet50-ibn)	ImageNet	84.2	61.5	15.7
ABD-Net	ImageNet	82.3	60.8	-