A Space Sim inspired data set to experiment with machine learning.
The objects of interest are cubes with a number on them. You may think of it as a 3D version of the MNIST data set.
In addition to the training samples, the data also contains pre-rendered flight through a 3D space scene.
The meta data for the flight is in flightpath/meta.pickle
. It contains, among
other things, the screen position of each cube in each frame. The
draw_bboxes.py
script shows how to use this information to draw bounding
boxes around all cubes.
Original Scene | After draw_bboxes.py |
---|---|
Download the dataset and use it in your ML model as you see fit.
The data set and code are available under the Apache Software License 2.0.
The numeral textures are royalty free images from PixaBay. The rendered images were created with PyHorde.
My own models are in the models/
folder. The results are documented
here
I would be happy to hear from anyone who uses it. Does it work for you? Is something missing? Which ML models have you built for it and how well did they do?