scipy or lap.lapjv? about ultralytics

[github-actions]

👋 Hello @flinzhao, thank you for your interest in Ultralytics YOLOv8 🚀! We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Join the vibrant Ultralytics Discord 🎧 community for real-time conversations and collaborations. This platform offers a perfect space to inquire, showcase your work, and connect with fellow Ultralytics users.

Install

Pip install the ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8.

pip install ultralytics

Environments

YOLOv8 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLOv8 Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

Hello @flinzhao,

Thank you for your detailed question and for providing a comprehensive test script! It's great to see such thorough investigation.

From your results, it appears that scipy.optimize.linear_sum_assignment is indeed faster than lap.lapjv for the cost matrices you tested. This aligns with some observations in the community, although the performance can vary depending on the specific use case and the size of the cost matrices.

The linear_assignment function in ultralytics/trackers/utils/matching.py defaults to using lap.lapjv because it can sometimes offer better performance on larger and more complex matrices, and it has been a reliable choice in various tracking scenarios. However, your findings suggest that for smaller matrices, scipy might be more efficient.

Here's a summary of your test results for quick reference:

Scipy (size 10): 0.000059 seconds
LAP (size 10): 0.000147 seconds
Scipy (size 50): 0.000103 seconds
LAP (size 50): 0.000213 seconds
Scipy (size 100): 0.000424 seconds
LAP (size 100): 0.000955 seconds
Scipy (size 200): 0.001190 seconds
LAP (size 200): 0.004794 seconds
Scipy (size 500): 0.006916 seconds
LAP (size 500): 0.028149 seconds

Given these results, you might consider setting use_lap=False in your specific application if you consistently observe better performance with scipy. However, it's important to test this in the context of your full application to ensure that the change does not negatively impact other aspects of performance or accuracy.

If you decide to make this change, you can adjust the linear_assignment function call in your code as follows:

matches, unmatched_a, unmatched_b = linear_assignment(cost_matrix, thresh, use_lap=False)

For further details on the linear_assignment function and other utilities, you can refer to the Ultralytics documentation on matching utilities.

If you encounter any issues or have further questions, feel free to reach out. We're here to help!