Dungeon Maps

Dungeon Maps is a powerful, lightweight PyTorch library for depth map manipulations, which is originally developed as a 2D mapping system for solving navigation problems: it can produce accurate 2D semantic top-down views from the depth map observations along with semantic segmentation predictions. See this Habitat object-goal navigation example:

output.mp4

Dungeon Maps also provides other functionalities, e.g. ego-centric motion flow calculation from depth maps. Scroll down to Demos to see more functionalities of this library.

This code is used by:

• HK Yang, TC Hsiao, et al. (2022). Investigation of Factorized Optical Flows as Mid-Level Representations. IROS 2022. Paper
• "Kemono", a rule-based object-goal navgiation agent for Habitat Challenge 2022. Ending2015a/kemono-habitat-2022

Version: 0.0.3a1

Features

	Batching	Multi-channels	GPU acceleration
Orthographic projection (Top-down map)	✔️	✔️	✔️
Egocentric motion flow	✔️	✔️	✔️
3D affine transformation (Camera space)	✔️	✔️	✔️
2D affine transformation (Top-down view)	✔️	✔️	✔️
Map builder	✔️	✔️	✔️

Installation

Basic requirements:

• Python >= 3.6
• PyTorch >= 1.8.0
• PyTorch Scatter

Install from pip

pip install dungeon_maps

Install from GitHub repo

pip install git+https://github.com.Ending2015a/dungeon_map.git@master

Install demos

pip install dungeon_maps[sim]

Demos

Orthographic projection

Depth maps

(Watch this video in high quality)

This example shows how to project depth maps to top-down maps.

• Top left: RGB
• Top right: Depth map
• Bottom left: top-down maps in local space
• Bottom right: top-down maps in global space

Run this example

python -m dungeon_maps.demos.height_map.run

Control: W, A, S, D. Q for exit

Semantic segmentations

(Watch this video in height quality)

This example shows how to project arbitrary value maps, e.g. semantic segmentation, to top-down maps.

• Top left: RGB
• Top center: Depth map
• Top right: Semantic segmentation
• Bottom left: top-down maps in local space
• Bottom right: top-down maps in global space

Run this example

python -m dungeon_maps.demos.object_map.run

Control: W, A, S, D. Q for exit

Ego-centric motion flow

(Watch this video in high quality)

This example shows how to calculate the flow fields caused by camera motion.

Run this example

python -m dungeon_maps.demos.ego_flow.run