Computer Vision News - February 2021

( ′ , ′ ) = ( + 1 ( ), + 2 ( )) RAFT is a new end-to-end trainable model dealing with the delicate task of optical flow . This refers to all those algorithms that have as ultimate scope to estimate per-pixel motion between video frames , which of course can become particularly difficult in real-life situations, where fast moving objects, occlusions, blurs and any kind of unpredictable changes are frequently found. As much challenging as it can be, optical flow is similarly useful. There is a wide range of computer vision tasks which could be substantially improved by a good estimation of the flow between different frames , or “instances in time”. Mathematically, the problem of optical flow can be expressed as: Finding a dense displacement field which maps each pixel in an image to the corresponding coordinates in the next image. Before the Deep Learning revolution , traditional optimization-based approaches were hand-crafted and used a continuous formulation of the above, based on the iterative refinement of a single estimate of optical flow. On the contrary, more ( , ) ( , ) 4 Research RAFT: Recurrent All-Pairs Field Transforms for Optical Flow Every month, Computer Vision News reviews a brilliant research paper. We started the new year with a great January issue focused on medical imaging research. Let’s just now have look at a completely different angle of computer vision, with a paper called RAFT: Recurrent All-Pairs Field Transforms for Optical Flow , written by a duo from Princeton (Zachary Teed and Jia Deng). We are indebted to the two authors for allowing us to use their images to illustrate this review. Their paper can be found at this link . by Marica Muffoletto BEST PAPER ECCV 2020