Semantic Segmentation

Simple inference implementation with trained HRNet on MIT ADE20K dataset, using PyTorch 1.6.0. Most of the code taken from [1]. Purpose of this project is to unify sky pixels with ultra high prediction confidence to a single color, in order to remove sun light effect and inconsistent cloud texture.

Usage

1. Download pretrained model from http://sceneparsing.csail.mit.edu/model/pytorch and store them in ./ade20k-hrnetv2-c1/.

2. Specify your test input image directory and test output image directory in semantic_segmentation.py.

   image_path = './input/'

   output_path = './output/'

3. Specify if you want to use GPU or not (-1 for cpu, others for gpu index) in semantic_segmentation.py.

   gpu = -1

4. Run python semantic_segmentation.py

5. All the inference results will be stored in your output_path

Results

semantic map

sky filtered image

Source

[1] Original code: https://github.com/CSAILVision/semantic-segmentation-pytorch.

[2] HRNet: https://arxiv.org/abs/1904.04514.