Geture Recognition

We trained a gesture recognition model based on VGG16 network. It can recognize four kinds of hand gestures. You can also train with sparsity, prune and finetune to get a tiny-model with slight accuarcy drop. The method comes from Learning Efficient Convolutional Networks through Network Slimming

Hand Gesture

Supported Hand Gesture as follow

train

python3 train.py --lr 0.1 --epochs 160 --weight-decay 5e-4 --mixup --alpha 0.2 

Train with sparsity

python3 train.py --lr 0.1 --epochs 160 --weight-decay 5e-4 --mixup --alpha 0.2 --sr 

Prune

python3 prune.py --model best_model_baseline.pth.tar --percent 0.7

Finetune

python3 train.py --refine best_model_pruned.pth.tar --lr 0.01 --epochs 60 --weight-decay 5e-4

Test

python3 test.py --model best_model_finetuned.pth.tar

Result

model	Accuarcy
baseline	94.4%
sparsity	93.6%
pruned	20%
finetuned	93%

Pretrained

pretrained weights
code: m73b

Demo

You can test it by using the code from https://github.com/lippman1125/YOLOv3-complete-pruning

python3 detect.py --cfg cfg/yolov3-hand.cfg --data data/oxfordhand.data --weights weights/best.pt --source 0 --gesture ../gesture_recog/model_best_finetune.pth.tar 

Reference

https://github.com/foolwood/pytorch-slimming