Awesome
DetNet : A Backbone network for Object Detection
This repository tries to reproduce the result in DetNet: A Backbone network for Object Detection using mxnet.
DetNet is evolved from fpn, thus one can change this repository to fpn easily.
demo result
set up environment
1.Clone this repository into the directory.
git clone git https://github.com/BigDeviltjj/mxnet-detnet.git
2.Download coco dataset into data directory.
3.Runsh init.sh
.
4.You may need to install python modules including cv2, matplotlib and numpy etc.
mAP
Model | Training data | Test data | mAP |
---|---|---|---|
detnet_coco-0016.params(multi_scale) | train2017 | val2017 | 39.7 |
demo
Put the images you want to detect in images/, then run
python demo.py
detected images will be saved in det_images/
train the model
You need to put the coco2017 dataset in data/ folder, and download pretrained model from detnet59-0000.params. Then put them in folder model/pretrained_mode, then run
python train_end2end.py
evaluate the model
Download the compressed trained model and symbol file detnet_coco-0016.params and unzip them then put them in folder output/detnet/coco/detnet/train2017/, then run
python test.py