The Garden of Forking Paths: Towards Multi-Future Trajectory Prediction

[Fujiry0]

Paper: http://openaccess.thecvf.com/content_CVPR_2020/papers/Liang_The_Garden_of_Forking_Paths_Towards_Multi-Future_Trajectory_Prediction_CVPR_2020_paper.pdf

Short summary: Predict the distribution over multiple possible future paths of people as they move through various visual scenes.

Details

• Method: Supervised deep neural networks
• Input: The ground truth location history, and a set of video frames, which are preprocessed by a semantic segmentation model.
• Output: Predicted locations of a single agent for multiple steps into the future.
• "Forking Paths" dataset
  • First 3D simulation dataset that is reconstructed from real-world scenarios complemented with a variety of human trajectory continuations for multi-future person trajectory prediction.
• New model "Multiverse"
  • The coarse location decoder's output: Heatmap over the 2D grid
  • The fine location decoder's output: Vector offset within each grid cell.
  • They use diverse beam search strategy to select the top K trajectory as inference
• SOTA on VIRAT/ActEV benchmark

Comment: This paper does not consider human to human interaction explicitly but they input semantic image in addition to location history.

GitHub : https://github.com/JunweiLiang/Multiverse Project Page: https://next.cs.cmu.edu/multiverse/

[jayakornv]

This work uses heatmap grids as main parametrization, then use beam-search to provide K candidate trajectories.