manuaero / Super-BPD

Super-BPD: Super Boundary-to-Pixel Direction for Fast Image Segmentation (CVPR 2020)

Super-BPD for Fast Image Segmentation (CVPR 2020)

Introduction

We propose direction-based super-BPD, an alternative to superpixel, for fast generic image segmentation, achieving state-of-the-art real-time result.

Prerequisite

• pytorch >= 1.3.0
• g++ 7

Dataset

• Download the BSDS500 & PascalContext Dataset, and unzip it into the Super-BPD/data folder.

Testing

• Compile cuda code for post-process.

cd post_process
python setup.py install

• Download the pre-trained PascalContext model and put it in the saved folder.

• Test the model and results will be saved in the test_pred_flux/PascalContext folder.

• SEISM is used for evaluation of image segmentation.

Training

• Download VGG-16 pretrained model.

python train.py --dataset PascalContext