In this paper, we considered the problem of detecting object take and release actions from untrimmed egocentric videos in an industrial domain. Rather than requiring that actions are recognized as they are observed, in an online fashion, we propose a quasi-online formulation in which take and release actions can be recognized shortly after they are observed, but keeping a low latency. We contribute a problem formulation, an evaluation protocol, and a baseline approach that relies on state-of-the-art components. Experiments on ENIGMA, a newly collected dataset of egocentric untrimmed videos of human-object interactions in an industrial scenario, and on THUMOS’14 show that the proposed approach achieves promising performance on quasi-online take/release action recognition and outperforms methods for online detection of action start on THUMOS’14 by $$+8.64\%$$ when an average latency of 2.19s is allowed. Code and supplementary material are available at https://github.com/fpv-iplab/Quasi-Online-Detection-Take-Release.
Quasi-Online Detection of Take and Release Actions from Egocentric Videos
Scavo R.;Ragusa F.;Farinella G. M.;Furnari A.
2023-01-01
Abstract
In this paper, we considered the problem of detecting object take and release actions from untrimmed egocentric videos in an industrial domain. Rather than requiring that actions are recognized as they are observed, in an online fashion, we propose a quasi-online formulation in which take and release actions can be recognized shortly after they are observed, but keeping a low latency. We contribute a problem formulation, an evaluation protocol, and a baseline approach that relies on state-of-the-art components. Experiments on ENIGMA, a newly collected dataset of egocentric untrimmed videos of human-object interactions in an industrial scenario, and on THUMOS’14 show that the proposed approach achieves promising performance on quasi-online take/release action recognition and outperforms methods for online detection of action start on THUMOS’14 by $$+8.64\%$$ when an average latency of 2.19s is allowed. Code and supplementary material are available at https://github.com/fpv-iplab/Quasi-Online-Detection-Take-Release.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.