Video object detection and segmentation

Abstract

We tackle the problem of realtime latency reduction in huge deep neural networks for the tasks of semantic segmentation and video object detection. Input dependent or conditional computation is employed to dynamically generate connectivity paths in a neural network for a given input. We also employ forecasting of unseen feature maps in cases where it is easier to generate such feature maps and avoid the expensive computation of backbone networks. Finally, we explore the utilization of a tracker and couple with a slow but accurate object detector using a reinforcement learning scheduling routine.