Model solves the pixel-semantic labeling task on the Cityscapes dataset. https://www.cityscapes-dataset.com/
Data: go to DOWNLOADS and start by downloading:
- Training, validation and test ground truth: gtFine_trainvaltest.zip (241MB)
- Training, validation and test images: leftImg8bit_trainvaltest.zip (11GB)
- save the cityscapes datasets in the
./datasets/citys
dir.
All pretrained models are based on a simple U-net architecture and stored in weights folder. The PyTorch implementation has been taken from this repo. Some changes are made and experimented with, see paper. If you want to run predictions use save_prediction.py and make sure the designated folder for the predictions exists.
Scores on the test set: link to Cityscapes benchmark
categories | IoU |
---|---|
construction | 66.7421 |
flat | 89.7289 |
human | 5.89255 |
nature | 74.5018 |
object | 3.63665 |
sky | 78.8369 |
vehicle | 48.5196 |
--------------- | ---------- |
Score Average | 52.5512 |