KidsGUARD: fine grained approach for child unsafe video representation and detection

Increasingly more and more videos are being uploaded on video sharing platforms, and a significant number of viewers on these platforms are children. At times, these videos have violent or sexually explicit scenes (referred as child unsafe) to catch children's attention. To evade moderation, malicious video uploaders typically limit the child unsafe content to only a few frames in the video. Hence, a fine-grained approach, referred as KidsGUARD¹, to detect sparsely present child unsafe content is required. Prior approaches to content moderation either flag the entire video as inappropriate or use hand-crafted features derived from video frames. In this work, we leverage Long Short Term Memory (LSTM) based autoencoder to learn effective video representations of video descriptors obtained from using VGG16 Convolutional Neural Network (CNN). Encoded video representations are fed into LSTM classifier for detection of sparse child unsafe video content. To evaluate this approach, we create a dataset of 109,835 video clips curated specifically for child unsafe content. We find that deep learning approach (1) detects fine-grained child unsafe video content with the granularity of 1 second, (2) identifies even sparsely location child unsafe video content by achieving a high recall of 81% at high precision of 80%, and (3) outperforms baseline video encoding approaches based on like Fisher Vector (FV) and Vector of Locally Aggregated Descriptors (VLAD).