ECCV 2022 - Wednesday

Future prediction is a critical planning requirement for self-driving, where it must consider the location and motion of all objects around a vehicle. Previous work has introduced the concept of fusing information from multiple cameras into a compact bird’s -eye view (BEV) representation to perform this prediction. However, with numerous possible futures needing to be covered, the quality of predictions can degrade over time. This work proposes a future segmentation method in a compact BEV representation , using time-dependent stochastic latent variables to model future uncertainty and the accuracy of predictions. “ Think about a case where you’re coming into a junction, and a car in front of you can turn right, or it can go straight ahead, ” Kaan poses. “ You need to know both these situations for planning. If you know the possibilities, you can plan ahead and act accordingly. There are multiple possible futures, and the prediction models must cover most of them for 22 Poster Presentation StretchBEV: Stretching Future Instance Prediction Spatially and Temporally Adil Kaan Akan ( right in the photo ) has just graduated from his Master’s at Koç University in Istanbul and is currently looking for PhD and internship opportunities. He works on stochastic future prediction and video understanding under the supervision of Fatma Güney ( left ), an Assistant Professor at Koç. They speak to us ahead of their poster session this afternoon.