A captivated audience attends a technical presentation on "Optimal Transport for Pseudo-labeling" at a conference. The projector screen displays a mathematical formulation aimed at optimizing the transport problem for \(N\) training images and \(K\) clusters or classes. The slide illustrates the optimization problem, including the minimization equation and the constraints on the transport matrix \(T\). A red box highlights a key part of the formulation. The clear message at the bottom of the slide emphasizes the requirement that all images must be labeled. The credits at the bottom mention Ming Xu and Stephen Gould for their work on "Temporally Consistent Unbalanced Optimal Transport for Unsupervised Video Segmentation," hinting at the detailed and advanced nature of the topic.
Text transcribed from the image:
Optimal Transport for Pseudo-labe
For N training images and K clusters/classes, solve
minimize
TER NxK
+
(C,T),
subject to
T1K = 1N,
T+¹Ñ = }¹K,
All images must be labelled
Ming Xu and Stephen Gould. Temporally Consistent Unbalanced Optimal Transport for Unsupervised
ת
Seg
n