Probabilistic Future Prediction for Video Scene Understanding
This addresses the critical need for robust future prediction in autonomous vehicles, though it builds incrementally on existing probabilistic and multi-task prediction methods.
The paper tackles the problem of predicting multiple aspects of future urban scenes from video for autonomous driving by jointly predicting ego-motion, static scenes, and dynamic agent motions probabilistically, achieving consistent future sampling from a compact latent space.
We present a novel deep learning architecture for probabilistic future prediction from video. We predict the future semantics, geometry and motion of complex real-world urban scenes and use this representation to control an autonomous vehicle. This work is the first to jointly predict ego-motion, static scene, and the motion of dynamic agents in a probabilistic manner, which allows sampling consistent, highly probable futures from a compact latent space. Our model learns a representation from RGB video with a spatio-temporal convolutional module. The learned representation can be explicitly decoded to future semantic segmentation, depth, and optical flow, in addition to being an input to a learnt driving policy. To model the stochasticity of the future, we introduce a conditional variational approach which minimises the divergence between the present distribution (what could happen given what we have seen) and the future distribution (what we observe actually happens). During inference, diverse futures are generated by sampling from the present distribution.