Learning-Deep-Learning

HOME: Heatmap Output for future Motion Estimation

August 2022

tl;dr: Using heatmap as output is a flexible way to balance MR and ADE.

Overall impression

The paper itself did not propose super innovative ideas. Instead it demonstrates a solid engineering work, and provides more insights into the advantages of the existing technical components.

The paper has two major contributions. First it uses CNN + attention in the backbone in parallel to encode heterogeneous input. Second it shows that the heatmap output (following the example of TNT) is a very flexible way to represent output as it allows different sampling strategy to tune the performance without retrain the model.

The challenge in behavior prediction is to avoid big failures where a possibility (a certain mode) is not considered at all. This has more impact to the safety of autonomous driving than having the absolute closes trajectory to the ground truth. –> This is the first principle we need to bear in mind when doing projects in behavior prediction. Multimodal is your friend, and mode collapse is your enemy.

In a way, HOME is a dense form of TNT and thus more like DenseTNT.

Key ideas

Technical details

Notes