Implementation of this paper for semantic segmentation. Written using tensorflow 2.0, and the model is implemented as a subclassed tf.keras.Model.
There are some differences between this implementation and the paper:
- Use normal resnet instead of wide resnet
- Use generalised dice loss instead of cross entropy for the edge segmentation
At the moment the best model I have trained is much worse than the paper, but this was because I did not replace the final downsampling layers in the resnet with atrous convolutions. I have done this now, but need to get access to a bunch of gpus to train on a decent batch size.
| Implementation | road | sidewalk | building | wall | fence | pole | traffic light | traffic sign | vegetation | terrain | sky | person | rider | car | truck | bus | train | motorcycle | bicycle |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Paper | 98.3 | 86.3 | 93.3 | 55.8 | 64 | 70.8 | 75.9 | 83.1 | 93 | 65.1 | 95.2 | 85.3 | 67.9 | 96 | 80.8 | 91.2 | 83.3 | 69.6 | 80.4 |
| This repo | 95.4 | 72.3 | 86.5 | 49.1 | 49 | 40.6 | 40.7 | 56.9 | 87.7 | 53.4 | 89 | 70.5 | 52.8 | 90.1 | 71.9 | 76.5 | 58.7 | 50 | 63.8 |
- Train better model on multiple gpus