For many years\, discrete optimization models such as cond itional random fields (CRFs) have defined the stat e-of-the-art for classical correspondence problems such as motion and stereo. One of the most import ant ingredients in those models is the choice of t he feature transform that is used to compute the s imilarity between images patches. For a long time \, hand crafted features such as the celebrated sc ale invariant feature transform (SIFT) defined the state-of-the-art. Triggered by the recent success of convolutional neural networks (CNNs)\, it is q uite natural to learn such a feature transform fro m data. In this talk\, I will show how to efficien tly learn such CNN features from data using an end -to-end learning approach. It turns out that our l earned models yields state-of-the-art results on a number of established benchmark databases.

Related Links

- https://arxi v.org/pdf/1707.06427 - Scalable Full Flow with Learned Binary Descriptors
- https://arxiv.org/pdf/1611.10229 - End-to-En d Training of Hybrid CNN-CRF Models for Stereo