Human-Activity-Recognition Temporal Segment Networks with RGB Diff produced 88% classification accuracy on validation data.