[한국인공지능학회] Redefining Temporal Information in Video Classification Problems - 함우주, 김한솔, 손재원, 김광수
- 인공지능융합 연구실
- 조회수219
- 2023-07-20
Abstract
The main challenge in machine-learning for video classification is understanding ‘Spatial In- formation’ and ‘Temporal Information.’ While significant progress has been made in extract- ing spatial information by developing 2D image classification models, the ‘Temporal Informa- tion’ extraction has not advanced as much. One possible reason is that a comprehensive defini- tion of temporal information has not yet been es- tablished. This paper proposes a novel definition of ‘Temporal Information’ in video classifica- tion consisting of ‘Movement Information’ and ‘Temporal Ordering Information.’ To demon- strate this, we conduct simple experiments using different timeline variations: Original, Reverse, and Stack. Furthermore, we evaluate how well- existing video classification models capture tem- poral information. To assess the meaningfulness of temporal ordering information in the feature vector obtained from video classification mod- els, we modify the classifier to predict the Origi- nal and Reverse data. These experiments show that most existing video classification models struggle to recognize temporal ordering informa- tion. Our findings are validated using benchmark datasets such as UCF101 and Kinetics400, along with several well-established baseline models.