Inference script and frozen inference graph with fine tuned weights for semantic segmentation on images from the KITTI dataset with TensorFlow. While the weights provided by the DeepLab authors reach an mIoU of 44% on the KITTI validation set, the fine-tuned weights reach an mIoU of 72.73 %.

Clone the repository: git clone https://github.com/hiwad-aziz/kitti_deeplab.git

Download the frozen inference graph: wget --no-check-certificate -r 'https://docs.google.com/uc?export=download&id=1ynVTP5CjMui2J3j2JLJDii8oC0ef_WpL' -O frozen_inference_graph.zip

Run the script: python inference.py path_to_frozen_inference_graph.pb path_to_image_folder/

https://github.com/tensorflow/models/tree/master/research/deeplab

http://www.cvlibs.net/datasets/kitti/