Motion Aware Data Sampling Using Sequential Frames for Deep Learning Models

In this study, an alternative data sampling method is proposed for deep learning models. In the suggested method, 3-stage sampling is done by using sequential frames. In the sampling process, a 5-minute 30 fps video is re-sampled to 5 fps and a total of 1500 frames are obtained. Optical flow and random sampling operations are performed sequentially on the frames obtained. In optical flow methods, 4 different thresholds are used while determining the magnitude vector. When performing random sampling, a total of 4 different data sets were created with 3 different sample rates. As a result of experiments with different sample rates and thresholds, a maximum of 1459 and a minimum of 757 frames are obtained from 1500 frames. The number of data obtained for each sub-data set is different. Obtained sub-data sets can be used for model generalization according to the needs of the environment in which the deep learning model will be used. To evaluate the performance of a deep learning model from different perspectives, experiments were carried out with the obtained data sets. As a result of the experiments, it was seen that the learning model cannot be directly associated with the number of data. According to different evaluation metrics, it has been evaluated which data set is more suitable for various environments.