Trajectory prediction using perception dataset

[next-gen-machine]

Can we directly use Waymo perception dataset for predicting trajectory of pedestrians ?

[scott-ettinger]

I think the answer depends on what you are trying to do. The perception dataset has a limited amount of data to use to train trajectory models compared to the motion dataset. The motion dataset has about 2 orders of magnitude more motion data.

Is your interest in using the additional sensor data in the perception dataset? If so, the latest motion dataset release includes a past history of lidar data in addition to the labeled perception objects.

[next-gen-machine]

Hii @scott-ettinger , Thanks for quick reply. I have two questions:

1) Can I use Waymo perception dataset in similar way as PIE or JAAD dataset for Pedestrian trajectory prediction ? because I want to implement it on Bi-Trap paper for trajectory prediction.

2) Since Waymo Motion dataset don't contain camera images regarding scenario, can I any how get camera images/videos regarding each scenario of Motion dataset ?

Because my aim is to create camera-based trajectory prediction model.

[scott-ettinger]

We do not currently supply camera images for the motion dataset. I don't have any information on how to compare with the two datasets you mentioned. The perception dataset does contain images and pedestrian labels, you would likely need to convert them to the format you need for your code.