Skip to content
BiSeNet based on pytorch
Branch: master
Clone or download
Latest commit bd993f7 May 21, 2019

README.md

BiSeNet

BiSeNet based on pytorch 0.4.1 and python 3.6

Dataset

Download CamVid dataset from Semantic-Segmentation-Suite
Thanks GeorgeSeif for his great job!

Pretrained model

Download epoch_295.pth in GoogleDrive and put it in ./checkpoints
The net structure has been modified, so the pretrained model does not match the current model.

Demo

python demo.py

Result

Original GT Predict

Train

python train.py

Use tensorboard to see the real-time loss and accuracy

loss on train

pixel precision on val

Test

python test.py

Result

Method Cropped Resized
Pixel Accuracy 94.1 93.2

Cropped and Resized means two image processing method to make the input image size fixed, it seems like Cropped input images get better result. I guess it's because cropped masks keep the original ground truth information while resized loss it.

Future work

  • Finish real-time segmentation with camera or pre-load video
  • Update pixel accuracy calculation algorithm

Reference

You can’t perform that action at this time.