A Quick Review and Performance Analysis of Custom and Transfer Learning CNN Architectures for Event Detection in Videos

Event/Action detection in videos is a growing research interest as it has numerous applications such as patient monitoring in health care, anomaly detection in surveillance systems, retrieval of video, human and computer interactions, gaming environment, entertainment environment etc. This is all because of one and only Deep learning due to its capability to outperform the conventional hand-crafted feature extraction algorithms. Transfer learning is significant in training deep neural networks with small datasets. The objective of this paper is to compare popular pretrained CNN models (in Top-1 accuracy) with custom CNN, built from scratch to detect human actions in UCF11 dataset.