This is the presentation for the final project of the course Intro to Computer Vision. Our program estimates the distance traveled by a basketball player based on video footage. We combined camera calibration, object tracking, depth estimation with deep learning, and image transformations in this project.
See more demo videos under Release
Python 3
pip install -r requirements.txtFor camera calibration, see the README inside the camera_calibration folder
For depth prediction, see the README inside the depth folder
-
Download our BallerSet dataset from the release page.
-
Unzip BallerSet and place the contents in the
./depthfolder -
Run our program with the command
python main.pyThe camera calibration has an average absolute percent error of 11.78%, the depth estimation has an average absolute percent error of 11.67%. Cumulatively, BallerStats has an average percent error of 46% with a standard deviation of 17% on our BallerSet dataset. The model always overestimates the distance traveled.
Robert Buckley, Richard Guan, Jensen Hwa, Tony Pan, Calvin Zheng
- BallerStats used open-source code from Robust Monocular Depth Estimation to generate depth disparity maps. We modified some of their code in accordance with the MIT License agreement, and encapsulated everything related to depth prediction in the
./depthfolder of this repo. Citation for their paper:
@article{Ranftl2020,
author = {Ren\'{e} Ranftl and Katrin Lasinger and David Hafner and Konrad Schindler and Vladlen Koltun},
title = {Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer},
journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
year = {2020},
}
- This is the final project for the course EECS 442 Intro to Computer Vision at the University of Michigan Winter 2021 semester taught by Justin Johnson and David Fouhey.
We have listed our software under the MIT License