Precise Temporal Action Detection for Fine-grained Actions

With the development of computer vision, existing temporal action detection algorithms made sense for coarse-grained scenes, while they did not work well in actual scenes due to the action similarity and the action boundary ambiguity. Based on perframe-I3D, we proposed a more precise network for fine-grained actions-perframe-TICDC, which mainly made improvements in enhancing the temporal convolution features and fusing local temporal features. For better performance, we adopt overlapping slices and skewness parameter estimation methods. Experimental results on 50 Salads Dataset showed that the perframe-TICDC outperformed previous methods significantly on the accuracy of temporal action boundary detection.