Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Decrease default threshold from 30 -> 27 #246

Closed
Breakthrough opened this issue Oct 12, 2021 · 5 comments
Closed

Decrease default threshold from 30 -> 27 #246

Breakthrough opened this issue Oct 12, 2021 · 5 comments
Milestone

Comments

@Breakthrough
Copy link
Owner

Breakthrough commented Oct 12, 2021

The default threshold (currently 30) should be lowered by 10% to achieve better accuracy for most inputs.

The current default was chosen as a nice round number, based on my own tests with a very small/limited dataset. After contacting the authors of https://videoprocessing.ai/benchmarks/shot-boundary-detection.html they kindly repeated the benchmark with threshold t = 27 (labelled "PyScene-v2" in the MSU Benchmarks) instead of 30. This scored much higher accuracy on the entire training set, which is fairly extensive.

tin2tin added a commit to tin2tin/Shot_Detection that referenced this issue Oct 16, 2021
@Breakthrough Breakthrough pinned this issue Oct 29, 2021
@Tetsujinfr
Copy link

Yes true, threshold=27 just worked better on most videos that I did test.
Thanks for the informed tip!

@CyberLykan
Copy link

What about regular threshold detection (not content)? Would the optimal value be 12 or something different?

@Breakthrough
Copy link
Owner Author

@CyberLykan good point, that was also just a guess based on some sample footage I created. The value could likely be raised higher to detect the fades earlier, but I think 12 is a pretty safe bet for most use cases. If anyone has a collection of samples that could be used as a heuristic though that would also be great.

@Tetsujinfr
Copy link

Fyi I had a video with lots of movements/dynamic camera motions where a threshold of 40 did produce good results while the 27 or 30 thresholds did not work well for me. So with the current approach, the threshold selection remains somewhat an art I think. This is a difficult problem to solve for properly and this repo is super useful anyway.

@Breakthrough
Copy link
Owner Author

@Tetsujinfr have you had an opportunity to test detect-adaptive? I'm curious if that works better for you, and if so, what parameters you end up using. Thanks!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

3 participants