FengMingquan-sjtu / DSPD

Decoupled Sub-Pixel Detection

1. Project description

Sub-pixel detection project in SJTU CS386. Our object is to detect interest points in sub-pixel accuracy. We proposed Decoupled Sub-Pixel Detection(DSPD), a novel formulation of Sub-Pixel Detection task.

The implementation is based on Python 3 and Pytorch 1.3.1. Experiments runs on Ubuntu and MacOS.

2. Usage

This part will run a simple demo of .

2.1. Install requirements

pip install -r requirements.txt

2.2. Prepare datasets

Put train HR(High-Resolution) images in ./data/intput/test_input_small We already put some demo images from DIV2K in the folder. If you want to use more images, DIV2K full dataset is available at https://github.com/xinntao/BasicSR/wiki/Prepare-datasets-in-LMDB-format .

2.3. Download pre-trained model

Please download SAN model from https://github.com/daitao/SAN, and EDSR, MDSR model from https://github.com/thstkdgus35/EDSR-PyTorch. And put these model files in ./weight.

2.4. Generate dataset

This step will split images into same shape, perform data augment and generate ground truth.

python utils/dataPrepare.py

2.5. Evaluate models

This step will run DSPD model, and print confusion matrix. Defaultly, it uses EDSR model, with scale factor = 2.

python eval.py

2.6. Check outputs

The SR images and detection images can be find in ./data/output.

3.Remarks

Current code is based on our previous End-to-End approach in https://github.com/FengMingquan-sjtu/spd.