Real-time computer vision on low-end boards via clustering motion vectors

David Denisov, Dan Feldman, Michael Segal, Shlomi Dolev

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Coreset, Motion vectors, Segments, Robotics, Structure from motion, non-convex optimization

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

In this work, we suggest computer vision methods, specifically for video tracking and map creation from video.
To this end, we utilize motion vectors and clusters, which are computed very efficiently in standard video encoders, usually via dedicated hardware.
We suggest a provably good tracking algorithm for clustering these vectors, by considering them as segments.
For this, we utilize a definition of a \emph{coreset} which is essentially a weighted set of points that approximates the fitting loss for every model, up to a multiplicative factor of $1\pm\varepsilon$.
Our method supports $M$-estimators that are robust to outliers, convex shapes, lines, and hyper-planes.
We demonstrate the empirical contribution of our clustering method for video tracking and map creation from video, by running it on micro-computers (Le-Potato and Raspberry Pi) on synthetic and real-world videos with real-time running time.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9453