February 2022
tl;dr: Predict diverse set of future targets and then use target to drive trajectory prediction.
The paper described the core drawbacks of previous methods, involving sampling latent states (VAE, GAN), or fixed anchors (coverNet, MultiPath).
TNT has the following advantages
- supervised training
- deterministic inference
- interpretable
- adaptive anchors
- likelihood estimation
The target, or final state capture most uncertainty of a trajectory. TNT decompose the distribution of futures by conditioning on targets, and then marginalizing over them.
The anchor-based method is improved by DenseTNT to be anchor-free, which also eliminated the NMS process by learning.
- Step 1: target prediction, based on manually chosen anchors
- Step 2: Motion estimation, conditioned on targets
- Step 3: Trajectory scoring/selection, with scoring and NMS
- Vectorized (sparse) encoding with VectorNet.