GTA-3D Dataset

A dataset of 2D imagery, 3D point cloud data, and 3D vehicle bounding box labels all generated using the Grand Theft Auto 5 game engine. The dataset contains image and depth map data captured at 1680x1050 resolution and oriented 3D bounding box labels of all vehicles. It is 55GB in total.

Created as part of my undergraduate thesis: 3D Object Localisation with Convolutional Neural Networks

alt text

Using the dataset

Helper classes to consume the data with are provided in gta.py. To read a single example use

from gta import GTAData

filename = 'data/gta_test/3fd50f7b-658b-4ef4-bb17-dfc1f287def8_00000819'
data = GTAData(filename)

img = data.load_rgb() # 1680 x 1050 x 3 ndarray of pixel color intensities
print(len(data.vehicles)) # Number of vehicles in the scene
print(data.vehicles[0].get_bbox_oriented_birdseye()) # Coordinates of the oriented birdseye bounding box of a vehicle in the scene
print(data.vehicles[0].get_bbox_2d()) # Coordinates of a bounding box in the image plane for a vehicle in the scene

An example 3D point cloud visualisation with bounding boxes is shown in test_vis.py