AUEB Stats Seminar by Andrea Mecca

ΚΥΚΛΟΣ ΣΕΜΙΝΑΡΙΩΝ ΣΤΑΤΙΣΤΙΚΗΣ 2026-2027

Andrea Mecca,
PhD Student, University of Florence, Florence, Italy, and Research Grant Holder, University of Siena, Siena, Italy
Statistical Knowledge Distillation for Tiny Object Detection in Tennis Videos Analysis

ΠΕΡΙΛΗΨΗ
In tennis video analysis, accurate ball detection is essential for match evaluation and performance analysis. Professional contexts often rely on costly high-resolution, high-frame-rate camera infrastructures, which are not available to amateurs. Deep-learning models such as TrackNet perform well in detecting small and fast-moving objects in sports such as tennis and badminton. However, their computational cost can hinder deployment in resource-constrained and real-time applications. We formulate knowledge distillation from a statistical perspective, treating pixel-level intermediate outputs from a TrackNet v1 teacher neural network as responses in a simpler and more transparent statistical student model. The resulting response variable is a semicontinuous random variable with a large clump at zero and continuous values in the open interval (0,1). Therefore, we adopt a hurdle-model framework, using logistic regression to model the probability of a positive response and beta regression for the conditional distribution of the response given that it is positive. For the logistic regression component, we propose a novel extension of a backfitting-type algorithm that jointly estimates the regression parameters and selects a specific covariate threshold in a data-driven manner using decision stumps. Since the joint analysis of multiple frames yields billions of pixel-level observations, full-data estimation is computationally prohibitive. Consequently, we also investigate informative subsampling techniques for logistic regression, with particular focus on adaptations of the local case-control subsampling method.