STDN: Scale-Transferrable Object Detection

A PyTorch Implementation of Scale-Transferrable Object Detection，the official code is not found,so I trained the model with pytorch,the code support:

• Support for the MS COCO dataset and VOC PASCAL dataset
• Support for stdn300,stdn321,stdn513 training and testing
• Support for mulltigpu training
• Support training and and testing in VOC and COCO

because my GPU is limited,so I only train the stdn300 in VOC0712,if your gpu is enough,you can train other model according to configs/*,the model can be downloaded in stdn300_densenet169。the map is 76.30,the map is lower because I have not pretrained model.

MAP in VOC2007

Original	Ours
78.1	76.30

Preparation

the supported version is pytorch-0.4.1 or pytorch-1.0

• tqdm
• opencv
• addict
• pytorch>=0.4

• Clone this repository.

git clone https://github.com/yxlijun/STDN.pytorch

• Compile the nms and coco tools:

sh make.sh

• Prepare dataset (e.g., VOC, COCO), refer to ssd.pytorch for detailed instructions.

train

you can train different set according to configs/*

python train.py --dataset VOC\COCO --config ./configs/stdn300_densenet169.py  

if you train with multi gpu

CUDA_VISIBLE_DEVICES=0,1 python train.py --dataset VOC\COCO --config ./configs/stdn300_densenet169.py --ngpu 2

eval

you can evaluate your model in voc and coco

python test.py --dataset VOC\COCO --trained_model ./weights/STDN_VOC_size300_netdensenet_epoch650.pth 

demo

you can test your image, First, download the pretrained stdn300_densenet169.pth file. Then, move the file to weights/.

python demo.py --dataset VOC\COCO --trained_model ./weights/STDN_VOC_size300_netdensenet_epoch650.pth --show  

You can see the image with drawed boxes as:

References

• Scale-Transferrable Object Detection
• M2Det