ICCV Daily 2019 - Wednesday

Anurag Ranjan is a PhD student at Max Planck Institute for Intelligent Systems with Honorary Professor Michael Black. He speaks to us following his poster session which we visited yesterday. The work is studying the vulnerabilities of optical flownetworks andcomparing them with classical methods and other vulnerable networks. Anurag says that in this case, they synthesized a small patch attack which was less than 1% of the image. When putting this patch in the scene and computing the optical flow using these networks, the optical flow failed miserably. The errors were about 400 per cent! These optical flow networks have activations when they’re not supposed to, with and without these attacks. This work investigates the problems with these neural networks to learn how to fix them. He says what’s really surprising is that the optical flow predictions from the network are a lot worse than he thought: “One of the biggest research questions is why this happens. We still don’t know exactly, but we proposed some heuristics of what might 6 Poster Presentation DA I L Y Attacking Optical Flow "This work investigates the problems with these neural networks to learn how to fix them."