31 lines (23 loc) · 1.39 KB

Data Card: ADE20k for Terrain Analysis

Dataset Overview

The underlying models (SegFormer B0/B2) are pre-trained on the ADE20k dataset, a scene parsing benchmark.

Source: MIT Computer Science and Artificial Intelligence Laboratory (CSAIL).
Scale: ~20k training images, ~2k validation images.
Classes: 150 fine-grained semantic categories.

Provenance

Authors: Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, Antonio Torralba.
Paper: "Scene Parsing through ADE20K Dataset", CVPR 2017.
URL: ADE20k Website

Class Mapping (Safe vs Hazard)

In this project, we map the 150 ADE20k classes to binary "Safe" / "Hazard" categories for navigation context.

Defaults

Safe:

grass, road, dirt, floor, path, vegetation, field, mountain, earth, plant

Hazard:

water, rock, sea, river, lake, pool, waterfall, boulder, cliff
person, vehicle, car, truck, bus, train, motorcycle, bicycle
snow, ice

Note: Any class not listed is treated as "Neutral" (Transparent).

Licensing & Terms

ADE20k License: Creative Commons or Research Use Only (Check usage terms on official site).
Usage: Since we use pre-trained weights from NVIDIA, we inherit the usage constraints of those weights which are derived from the dataset licensing.