Bird-eye's view for CARLA

Freeway-oriented shape	Learning By Cheating-like shape	Centerlines layer

(GIFs above present feature maps after applying as_rgb() function)

Motivation

During our research we found a very inspiring paper called Learning By Cheating. Bird-eye's view is made specifically to learn faster thanks to much simpler, 2D world representation (cheating oracle) which we think fits well in Reinforcement Learning setup.

This repository is an almost complete reimplementation that gives better performance and compatibility with most recent versions of CARLA. You can use it out-of-the-box as input for your model, and if necessary convert and visualize into RGB.

Features

• one-hot 3D feature map (8x2D layers, each representing other entities, e.g. road layer, pedestrians layer) - made specifically to feed your CNN
• feature map can be converted to an RGB image
• layers can be easily removed
• caching mechanism for static layers like: roads and lanes
• using OpenCV rendering (efficient, multi-threading friendly) instead of slow Pygame method
• huge FPS speedup thanks to restricted rendering (only agent's surroundings, not whole map)
• all CARLA maps are supported out-of-the-box, custom maps with valid OpenDrive file made in RoadRunner are also supported
• current implementation is specifically adjusted for highway scenarios (prolonged shape), but other shapes and crops are easy to implement

Installation

pip install carla-birdeye-view

How to run

Make sure that PYTHONPATH env variable contains CARLA distribution egg, so that carla package can be imported.

# Launch server instance
./CarlaUE4.sh

# (optional) For CARLA 0.9.8+ you may get additional performance improvement with this
python PythonAPI/util/config.py --no-rendering

# Preview while cruising on autopilot (birdview/__main__.py)
python -m carla_birdeye_view

Basic code usage

from carla_birdeye_view import BirdViewProducer, BirdViewCropType, PixelDimensions

birdview_producer = BirdViewProducer(
    client,  # carla.Client
    target_size=PixelDimensions(width=150, height=336),
    pixels_per_meter=4,
    crop_type=BirdViewCropType.FRONT_AND_REAR_AREA
)

# Input for your model - call it every simulation step
# returned result is np.ndarray with ones and zeros of shape (8, height, width)
birdview = birdview_producer.produce(
    agent_vehicle=agent  # carla.Actor (spawned vehicle)
)

# Use only if you want to visualize
# produces np.ndarray of shape (height, width, 3)
rgb = BirdViewProducer.as_rgb(birdview)

Contribution and feedback

We'd ❤️ to collct any feedback, issues and pull requests!

Credits

Project born at deepsense.ai, made by:

Michał Martyniak (@micmarty)