zhaowujie / EFPN-pytorch

EFPN

This repo and account are part of the supplementary material for an anonymous submission, so we do not accept any PR or issue currently.

Requirements

This code has been developed under Python3.6, PyTorch 0.3.1 and CUDA 9.0 on Ubuntu 16.04.

The python packages can be installed with

pip3 install -r requirements.txt

Compilation

Compile the CUDA code:

cd lib  # please change to this directory
sh make.sh

CUDA_PATH defaults to /usr/loca/cuda. If you want to use a CUDA library on different path, change this line in lib/make.sh accordingly.

Pretrainded Models

We provide pretrainded models on Tsinghua-Tencent 100K and COCO for inference. The models are both based on Faster R-CNN with ResNeXt-101-EFPN.

Download our trained models from GoogleDrive or BaiduYun(Key:l24p), and put them into {repo_root}/checkpoints.

Inference

At present, the code only supports single GPU inference. You can specify the GPU id on line in tools/infer_simple.py (default set to GPU 0).

To visualize examples of Tsinghua-Tencent 100K, run with:

python tools/infer_simple.py --dataset tt100k --cfg configs/EFPN_X-101_TT100K.yaml --load_ckpt checkpoints/EFPN_X101_TT100K.pth --image_dir examples/tt100k --output_dir examples/res_tt100k

To visualize examples of Tsinghua-Tencent 100K, run with:

python tools/infer_simple.py --dataset coco2017 --cfg configs/EFPN_X-101_COCO.yaml --load_ckpt checkpoints/EFPN_X101_COCO.pth --image_dir examples/coco --output_dir examples/res_coco

and the detection results will be saved in examples/res_tt100k and examples/res_coco.

Examples

We provide several detection examples of EFPN in ./examples. You can also directly view them.

Training

TBA