Awesome
BiSeNet
BiSeNet based on pytorch 0.4.1 and python 3.6
Dataset
Download CamVid dataset from Google Drive or Baidu Yun(6xw4).
Pretrained model
Download best_dice_loss_miou_0.655.pth in Google Drive or in Baidu Yun(6y3e) and put it in ./checkpoints
Demo
python demo.py
Result
| Original | GT | Predict |
|---|---|---|
| <img src="https://github.com/ooooverflow/BiSeNet/blob/master/test.png" width="300" height="225" alt=""/> | <img src="https://github.com/ooooverflow/BiSeNet/blob/master/test_label.png" width="300" height="225" alt=""/> | <img src="https://github.com/ooooverflow/BiSeNet/blob/master/demo.png" width="300" height="225" alt=""/> |
Train
python train.py
Use tensorboard to see the real-time loss and accuracy
loss on train
<img src="https://github.com/ooooverflow/BiSeNet/blob/master/tfboard_loss.jpg" width="1343" height="260" alt=""/>pixel precision on val
<img src="https://github.com/ooooverflow/BiSeNet/blob/master/tfboard_precision.jpg" width="1343" height="260" alt=""/>miou on val
<img src="https://github.com/ooooverflow/BiSeNet/blob/master/tfboard_miou.jpg" width="1343" height="260" alt=""/>Test
python test.py
Result
| class | Bicyclist | Building | Car | Pole | Fence | Pedestrian | Road | Sidewalk | SignSymbol | Sky | Tree | miou |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| iou | 0.61 | 0.80 | 0.86 | 0.35 | 0.37 | 0.59 | 0.88 | 0.81 | 0.28 | 0.91 | 0.73 | 0.655 |
This time I train the model with dice loss and get better result than cross entropy loss. I did not use lots special training strategy, you can get much better result than this repo if using task-specific strategy.
This repo is mainly for proving the effeciveness of the model.
I also tried some simplified version of bisenet but it seems does not preform very well in CamVid dataset.
Speed
| Method | 640×320 | 1280×720 | 1920×1080 |
|---|---|---|---|
| Paper | 129.4 | 47.9 | 23 |
| This Repo | 126.8 | 53.7 | 23.6 |
This shows the speed comparison between paper and my implementation.
- The number in first row means input image resolution.
- The number in second and third row means FPS.
- The result is based on resnet-18.
Future work
- Finish real-time segmentation with camera or pre-load video