Deep Learning and its Relationship with Time (Talk)
In this talk I am going to present the work we have been doing at the Computer Vision Lab of the Technical University of Munich which started as an attempt to better deal with videos (and therefore the time domain) within neural network architectures.
Oddly enough, we ended up not including time at all in our proposed solutions. In the first work, we tackle the task of semi-supervised video object segmentation, i.e., the separation of an object from the background in a video, given the mask of the first frame. I will present One-Shot Video Object Segmentation (OSVOS), based on a fully-convolutional neural network architecture that is able to successively transfer generic semantic information, learned on ImageNet, to the task of foreground segmentation, and finally to learning the appearance of a single annotated object of the test sequence (hence one-shot). OSVOS is fast and improves the state of the art by a significant margin (79.8% vs 68.0%).
The second work I will present is a new CNN+LSTM architecture for camera pose regression for indoor and outdoor scenes. Contrary to most works, we make use of LSTM units on the CNN output in spatial coordinates in order to capture contextual information. This substantially enlarges the receptive field of each pixel leading to drastic improvements in localization performance.
I will also present a new large-scale indoor dataset with accurate ground truth from a laser scanner.
