Ryota Hashiguchi

Aoi Otani, Ryota Hashiguchi, Kazuki Omi et al.

In the design of action recognition models, the quality of videos is an important issue; however, the trade-off between the quality and performance is often ignored. In general, action recognition models are trained on high-quality videos, hence it is not known how the model performance degrades when tested on low-quality videos, and how much the quality of training videos affects the performance. The issue of video quality is important, however, it has not been studied so far. The goal of this study is to show the trade-off between the performance and the quality of training and test videos by quantitative performance evaluation of several action recognition models for transcoded videos in different qualities. First, we show how the video quality affects the performance of pre-trained models. We transcode the original validation videos of Kinetics400 by changing quality control parameters of JPEG (compression strength) and H.264/AVC (CRF). Then we use the transcoded videos to validate the pre-trained models. Second, we show how the models perform when trained on transcoded videos. We transcode the original training videos of Kinetics400 by changing the quality parameters of JPEG and H.264/AVC. Then we train the models on the transcoded training videos and validate them with the original and transcoded validation videos. Experimental results with JPEG transcoding show that there is no severe performance degradation (up to -1.5%) for compression strength smaller than 70 where no quality degradation is visually observed, and for larger than 80 the performance degrades linearly with respect to the quality index. Experiments with H.264/AVC transcoding show that there is no significant performance loss (up to -1%) with CRF30 while the total size of video files is reduced to 30%.

Ryota Hashiguchi, Toru Tamaki

Feature shifts have been shown to be useful for action recognition with CNN-based models since Temporal Shift Module (TSM) was proposed. It is based on frame-wise feature extraction with late fusion, and layer features are shifted along the time direction for the temporal interaction. TokenShift, a recent model based on Vision Transformer (ViT), also uses the temporal feature shift mechanism, which, however, does not fully exploit the structure of Multi-head Self-Attention (MSA) in ViT. In this paper, we propose Multi-head Self/Cross-Attention (MSCA), which fully utilizes the attention structure. TokenShift is based on a frame-wise ViT with features temporally shifted with successive frames (at time t+1 and t-1). In contrast, the proposed MSCA replaces MSA in the frame-wise ViT, and some MSA heads attend to successive frames instead of the current frame. The computation cost is the same as the frame-wise ViT and TokenShift as it simply changes the target to which the attention is taken. There is a choice about which of key, query, and value are taken from the successive frames, then we experimentally compared these variants with Kinetics400. We also investigate other variants in which the proposed MSCA is used along the patch dimension of ViT, instead of the head dimension. Experimental results show that a variant, MSCA-KV, shows the best performance and is better than TokenShift by 0.1% and then ViT by 1.2%.