This repository contains the implementation details of our paper: [arXiv:1809.09077]
"Incorporating Luminance, Depth and Color Information by a Fusion-based Networks for Semantic Segmentation"
by Shang-Wei Hung , Shao-Yuan Lo, Hsueh-Ming Hang

• Python 3
• Pytorch 0.4.1

LDFNet adopts two distinctive sub-networks in a parallel manner to process multiple information.
Also, LDFNet employs the luminance information to assist the processing of the depth information in the D&Y encoder.

LDFNet achieves a mIoU scores of 71.3 % on the Cityscapes dataset without any pretrained model.[Benchmarks]
For the resolution 512x1024 input, LDFNet can run at the speed of 18.4 and 27.7FPS on a singel Titan X and GTX 1080 Ti, respectively.

Flow the steps presented below:

• Download "train" and "model" files.
• Create your own global file and put "train" and "model" in it.
• Set the global file path in the very bottom of "main.py" in "train" file.
• Create "save" file in your global file, and the trained model and the validation results will be saved in the file you specify.

If you feel our LDFNet is useful for your research, please consider citing our paper:

• S.-W. Hung, S.-Y. Lo and H.-M. Hang, “Incorporating Luminance, Depth and Color Information by a Fusion-based Network for Semantic Segmentation,” in International Conference on Image Processing, 2019.