Computer Vision News - December 2019

Poster Presentation 26 Chiho Choi is a scientist at Honda Research Institute USA. He spoke to us following his poster session. Chiho’s work is about a relation-aware framework for future trajectory forecast which uses a set of images to extract relational behavior between road users and their surrounding environments. It uses images as input and tries to model both human-human interactions together with human-space interactions, as well as relational behavior between agents. In this domain, they are the first to use those relational behaviors for future trajectory forecast, but the real novelty is the specific relation gate module in their work. They are predicting the future, which is usually unpredictable, so there are lots of challenges to consider, including solve this by applying an uncertainty modeling method. Chiho explains: “In the literature there exists a Monte Carlo dropout method to model uncertainty of the neural network which is applied for semantic segmentation, image classification, and those kinds of things. We extend the use of Monte Carlo dropout for the prediction problem. However, it is very hard to say that our work achieves multimodality over the future prediction problembecause even usingMonte Carlo dropout, the variation of the predictions is very low. We would extend this work to incorporate actual multimodality to get rid of the uncertainty estimate.” Continuing to think about what’s next, Looking to Relations for Future Trajectory Forecast Best of ICCV 2019