Seg-PyTorch-Dev

This repository contains simple baseline for segmentation projects.

Structure of the repository:

• [additional]
  Some scripts and tools to prepare certain datasets
• [models]
  Segmentation models and functions for them (checkpoint saving, exporting, ..)
  There will also be custom loss functions
• [data]
  All for handling the data: pytorch datasets, samplers and functions for data processing
• train.py
  Simple script for training

How to use

1. Download dataset (works only with ADE20K for now) and run additional/ade20k_1csv_list.py, then additional/ade20k_2csv_sets.py. Don't forget to change parameters (paths) in these scripts!
2. Train with train.py.

There are no scripts for inference yet :( I need some time...

Datasets

• ADE20K [http://sceneparsing.csail.mit.edu/, scene Parsing train/val data]
  Scene-centric indoor and outdoor images.
• The Oxford-IIIT Pet [https://www.kaggle.com/devdgohil/the-oxfordiiit-pet-dataset]
  a 37 category pet dataset with roughly 100 images for each class created by the Visual Geometry Group at Oxford.
• [TODO] Makeup. Pixel Perfect Lips Segmentation [https://www.kaggle.com/olekslu/makeup-lips-segmentation-28k-samples]
  The data was gathered and annotated with the custom semisupervised image annotation algorithm.
• [TODO] Segmentation Full Body MADS Dataset [https://www.kaggle.com/tapakah68/segmentation-full-body-mads-dataset]
  The MADS dataset with segmented people, 1192 images.

Notes and TODOs

• download weights for torchvision models and organize their loading to models
• reorganize train.py script - all input data and parameters should be taken from config file
• test UNet model and add some other variations