Samera: a scalable and memory-efficient feature extraction algorithm for short 3D video segments

Tele-immersive systems, are growing in popularity and sophistication. They generate 3D video content in large scale, yielding challenges for executing data-mining tasks. Some of the tasks include classification of actions, recognizing and learning actor movements and so on. Fundamentally, these tasks require tagging and identifying of the features present in the tele-immersive 3D videos. We target the problem of 3D feature extraction, a relatively unexplored direction. In this paper we propose Samera, a scalable and memory-efficient feature extraction algorithm which works on short 3D video segments. The focus is on relevant portions of each frame, then uses a flow based technique across frames (in a short video segment) to extract features. Finally it is scalable, by representing the constructed feature vector as a binary vector using Bloom Filters. The results obtained from experiments performed on 3D video segments obtained from Laban Movement Analysis (LMA) show that the compression ratio achieved in Samera is 147.5 as compared to the original 3D videos.