readme.md

Vision Base - Segmentation

This repo use vision_base library extracts from FSNet to accelerate the development of segmentation related research.

The vision_base library contains generic codes for:

1. Config-based model/pipeline construction.
2. Common tools for pipeline organization, experiment logs.
3. Common mathematical, pytorch, numerical tools.

Recommended Practice

1. Write code for dataset fetching, dataset evaluator, and network model.
2. Adapt config files based on examples to launch dataset/network/training with your own model.
3. Use existing scripts and launchers to start experiments.
4. Minimize the modification in vision_base. Just write new classes in another folders if needed.

Model Introduction

This segmentation package implements a simple UNet with a Transformer block to conduct semantic segmentation on KITTI360 dataset.

Data Preparation

Download the KITTI360 data2D image sequences and data2D 2D annotations. Organized in {...}/KITTI-360/{data_2d_raw|data_2d_semantics}

After modifying paths in the configuration file. Launch training with the standard script.