Are current long-term video understanding datasets long-term?
This work addresses a critical issue for researchers in video understanding by showing that current datasets might not properly evaluate long-term action recognition models, potentially leading to misleading benchmarks.
The authors evaluated whether existing long-term video action recognition datasets truly require long-term information by defining a method to exclude videos solvable with short-term cues, testing it on Breakfast, CrossTask, and LVU datasets. They found these datasets can be solved using shortcuts based on short-term information, revealing they may not effectively assess long-term recognition.
Many real-world applications, from sport analysis to surveillance, benefit from automatic long-term action recognition. In the current deep learning paradigm for automatic action recognition, it is imperative that models are trained and tested on datasets and tasks that evaluate if such models actually learn and reason over long-term information. In this work, we propose a method to evaluate how suitable a video dataset is to evaluate models for long-term action recognition. To this end, we define a long-term action as excluding all the videos that can be correctly recognized using solely short-term information. We test this definition on existing long-term classification tasks on three popular real-world datasets, namely Breakfast, CrossTask and LVU, to determine if these datasets are truly evaluating long-term recognition. Our study reveals that these datasets can be effectively solved using shortcuts based on short-term information. Following this finding, we encourage long-term action recognition researchers to make use of datasets that need long-term information to be solved.