Simple-RFCN-PyTorch

A simple and concise implementation of the RFCN is given.

This project can be run with Pytorch 1.7.

Results

1. train on voc2007 & no OHEM 2. train on voc07+12 & OHEM

	Train on voc2007	Train on voc07+12
From scratch without OHEM	71.6%	/
From scratch with OHEM	72.5%	76.8%

Features

• The results are comparable with those described in the paper(RFCN).
• Very low cuda memory usage (about 3GB(training) and 1.7GB(testing) for resnet101).
• It can be run as pure Python code, no more build affair.

Requirements:

matplotlib==3.2.2
tqdm==4.47.0
numpy==1.18.5
visdom==0.1.8.9
fire==0.3.1
torchnet==0.0.4
opencv_contrib_python==4.5.1.48
scikit_image==0.16.2
torchvision==0.8.1
torch==1.7.0
cupy==8.4.0
Pillow==8.1.0
skimage==0.0

Usage

cd [RFCN-pytorch root_dir]

Train:

python -m visdom.server
python train.py RFCN_train

Access 'http://localhost:8097/' to view loss and mAP (real-time).

Eval:

python train.py RFCN_eval --load_path='checkPoints/rfcn_voc07_0.725_ohem.pth' --test_num=5000

Predict
Place the pictures to be predicted in predict/imgs folder.
Run command in terminal:

python predict.py predict --load_path='checkPoints/rfcn_voc07_0.725_ohem.pth'

Weights & CheckPoints

You can download the weights of ResNet101 and place it in weights folder.

You can download the pretrained model from Google Drive or 百度云盘(passwd: 9o15) and place it in checkPoints folder.

Acknowledgement

This project is writen by elbert-xiao, and thanks to the provider chenyuntc for the project simple-faster-rcnn-pytorch!

If you have any question, please feel free to open an issue.