Part III: Deep Learning vs SLAM
SLAM group discussion is really fun. Before we go into the important "deep learning vs slam" "discussion, I should say that every seminar contributor agrees: Semantics are necessary to build a larger and better SLAM system. There are lots of interesting little conversations about the future direction. In the debate, Marc Pollefeys (a well-known SfM and multi-view geometry researcher) reminds everyone that "robots are a killer application of SLAM" and advises us to stay focused on the "jackpot". This is very surprising, because SLAM is traditionally suitable for robotic problems, but robots have not been successful in the last few decades (Google Robotics?). ), leading to SLAM's focus shifted from robots to large-scale map-building (including Google Maps) and augmented reality. No one talked about robots at the seminar.
1. Integrating semantic information into SLAM
There is great interest in integrating semantics into today's best SLAM systems. When it comes to semantics, the SLAM community is unfortunately stuck in the world of visual word bags (bags-of-visual-words), and there is little New thought on how to integrate semantic information into their systems . At one end of the semantic, we have now seen a lot of real-time semantic segmentation demonstrations on CVPR/ICCV/ECCV (based on convolutional neural networks); In my opinion, SLAM needs deep learning, and deep learning needs SLAM as well.
Deep Learning vs SLAM