26/01/2026 2:41 PM - PACES
Problem
Deciding which features are good for tracking over multiple frames in a video.
Approach
They suggest a two part approach:
First, they define a good feature point. They do so by defining a mathematical criteria for what it means to be good for tracking and maximizing it. They argue that to be good for tracking means to be highly different than the surrounding pixels in squared difference. This definition leads to an objective of maximizing the minimum eigenvalue of the structure tensor (although they do not call it that yet). They note that really they also need to also put a cap on the conditioning number of the matrix, but due to the structure of the image the maximum value of an eigenvalue is bounded so just putting a lower bound on what either eigenvalue can be works as well.
Second, they define a good feature for tracking. They argue that when tracking over multiple frames, a good feature is one where the visual on the Nth frame is well approximated as an affine transformation from the 1st frame. This naturally rejects features like specular highlights and features made by occlusions and can also catch aliasing problems as those tend to jump around in value using this method.
Contribution
They seem to have introduced the notion of using the eigenvalues of the structure tensor as the criterion for a good feature point as well as proposing the use of an affine transform to check whether a feature point is still good to use over multiple frames.
Evaluation
They evaluate in multiple ways:
1. Proofs: They prove the optimality of the structure tensor given the assumption that a good feature is one that is easily differentiable from surrounding pixels.
2. Simulation: They simulate sequences of images of with known transformations between them and check whether they can recover the parameters using their method.
3. Real world evaluation: They then apply their method to a real video (while also seeminly introducing a new criteria which reject feature points if they overlap) and point out where their criteria reject bad feature points.
Substantiation
This is a strong evaluation as they prove both mathematically and empirically that their method works. I would have liked to see many more real world examples and comparisons to prior works, but as with previous weeks it seems that was not the fashion at the time.
27/01/2026 11:10 AM