This repo mainly based on jwyang/fpn.pytorch and jwyang/faster-rcnn.pytorch. I combine some code to let it ables to work in pytorch1.0 framework and get a more 75.8mAP(higher than faster rcnn & fpn0.4 repo) when training pascal voc 2007.

Iherent from them, this repo support multi GPU training, GPU version NMS and ROI Align pooling. Thanks a lot for jwyang. More usage introduction can be found in the upper two repo.

I benchmark this code thoroughly on pascal voc2007 (voc0712 is on the way). Below are the results:

1). PASCAL VOC 2007 (Train/Test: 07trainval/07test, scale=600, ROI Align，

clone the code

git clone https://github.com/tianyolanda/fpn_toyotakyushu.git

Then, create a folder:

cd fpn-pytorch1.0 
mkdir data
mkdir logs

The environment I run this code is under:

• Python 3.7
• Pytorch 1.0
• CUDA 10.0

visdom are support for visilization of loss curve

• VOC2007 or VOC07+12: Please follow the instructions in py-faster-rcnn to prepare VOC datasets. Actually, you can refer to any others. After downloading the data, creat softlinks in the folder data/.

• COCO dataset is not supported in this repo yet

Pretrained Model we need for FPN is ResNet101.

Download from here jwyang/faster-rcnn.pytorch

Download it and put it into the data/pretrained_model/.

Install all the python dependencies using pip:

pip install -r requirements.txt

Compile the cuda dependencies using following simple commands:

cd lib
python setup.py build develop

It will compile all the modules you need, including NMS, ROI_Align. Please check to compiled with coorsponding python version.

train voc2007:

CUDA_VISIBLE_DEVICES=0 python trainval_net.py --dataset pascal_voc --cuda

test voc2007:

CUDA_VISIBLE_DEVICES=0 python test_net.py --dataset pascal_voc --checksession 1 --checkepoch 12 --checkpoint 10021 --cuda

train voc07+12:

CUDA_VISIBLE_DEVICES=0 python trainval_net.py --dataset pascal_voc_0712 --cuda

Here I provide my trained FPN model(trained on pascal voc 2007, can detect 20 kinds of objects), you can simply test it without training.

Download the model from baiduyun

put the trained FPN model in /fpn-pytorch1.0/models/res101/pascal_voc/fpn_1_12_10021.pth

Put images to be detected in /fpn-pytorch1.0/images/

Then run demo.py

CUDA_VISIBLE_DEVICES=0 python demo.py --dataset pascal_voc --checkepoch 12 --cuda

• Train and test on VOC0712
• Train and test on COCO
• Support softNMS
• Support DetNet