Regularizing Oversmoothing of Temporal Convolutional Networks for Action Segmentation into Human Assembly Operations

• DOI: 10.2299/jsp.27.75
• Publication Date: 2023-06-30T22:17:50Z
• Authors: Keisuke Nakamura; Yoshitaka Yamamoto; Masafumi Nishimura; Yuki Shiono; Reiki Shirasawa; Takayuki Nakano; Takahiro Aoki
• Keywords: Normalization; Benchmark (surveying); Action Recognition

We investigated a (supervised) deep learning model for the automatic segmentation of manufacturing video data into sequence assembly operations. In action general human behavior, various temporal convolutional network (TCN)-based methods have been proposed and demonstrated to stable performance using features from captured extensive fields image frames. However, they often make it difficult detect unusual actions occurring in short durations, such as skipping this paper, we address drawback existing TCN by introducing two techniques differentiable group normalization jumping knowledge relax oversmoothing effect. Then, empirically show variations benchmark operational work segmentation. also present preliminary result on irregular detection.