Spatiotemporal Residual Networks for Video Action Recognition