Task Graph Maximum Likelihood Estimation for Procedural Activity Understanding in Egocentric Videos

TL;DR

This paper presents a gradient-based method for learning task graphs from procedural activities in egocentric videos, significantly improving understanding accuracy and enabling neural integration.

Contribution

It introduces a novel maximum likelihood optimization approach for task graph learning, surpassing hand-crafted methods and demonstrating strong results on multiple benchmarks.

Findings

Achieved +14.5% F1-score on CaptainCook4D

Improved online mistake detection by +19.8% on Assembly101-O

Top performance on Ego-Exo4D benchmark

Abstract

We introduce a gradient-based approach for learning task graphs from procedural activities, improving over hand-crafted methods. Our method directly optimizes edge weights via maximum likelihood, enabling integration into neural architectures. We validate our approach on CaptainCook4D, EgoPER, and EgoProceL, achieving +14.5%, +10.2%, and +13.6% F1-score improvements. Our feature-based approach for predicting task graphs from textual/video embeddings demonstrates emerging video understanding abilities. We also achieved top performance on the procedure understanding benchmark on Ego-Exo4D and significantly improved online mistake detection (+19.8% on Assembly101-O, +6.4% on EPIC-Tent-O). Code: https://github.com/fpv-iplab/Differentiable-Task-Graph-Learning.