This repository contains the source code to GUTS, as well as our reimplementation of UTS. Because these methods share so many components, they share one codebase. Instances of the Options class specify which method is used, including possible superpositions between the two methods, such as running UTS with Mask R-CNN as its detector.
Click here to see an example video of GUTS running on the UTOCS dataset
If you find this project useful, please cite our paper: https://ieeexplore.ieee.org/abstract/document/9921949
This method can be used alongside the UTOCS dataset, in addition to custom datasets. To run on your own data, you can either create a dataset object similar to the UTOCS class in utocs.py, or look at guts.py and test_guts.py and apply the method to images, ground surfaces and camera calibration matrices you load manually.
Both GUTS and our implementation of UTS makes the following assumptions on your data:
- The video consist of RGB images at a reasonable resolution like 720p or 1080p
- The camera is fully calibrated with a 4x3 projection matrix, and radial distortion is removed from the images.
- Some sampled ground points in world coordinates exists that are visible in the scene. For UTS, these are only used to define a plane perpendicular to the z-axis, at the z-level that is the average of the points.
- GUTS supports cars, trucks, buses, pedestrians and bicyclists. UTS supports cars, trucks and buses.
You will need the following dependencies:
- A powerful, modern Linux computer with an Nvidia GPU and correctly installed drivers, including CUDA 11.4
gitsingularity(we use version 3.8.1)
git clone https://github.com/ahrnbom/guts
cd guts
./run_singularity.sh bash
# Inside the Singularity image now
python -m pytest .
When inside the Singularity container, press Ctrl+D to exit to your own shell.
The tests should all pass, albeit with some warnings. After running the tests, inspect the video in output/demo.mp4 and see that tracking is performed reasonably well.