Prediction Challenge Dataset


I’m new to nuscenes and I’d like to try to participate in the prediction challenge, however, before that I’ve got couple of following questions that needs answering:

  1. In the section “General rules” in the first bullet point you mention " We have created a hold out set for validation from the training set called the train_val set", now do I just download the whole dataset as normal and split the train_val using ‘create_splits_scenes’ from, or is there a separate dataset for that actual task?

  2. Does the prediction involve predicting motion for ego vehicle or for annotated agents?

  3. If the prediction involves predicting motion for annotated agents, does the data from annotation e.g. translation, rotation, corresponds to true values of the actual agent?



  1. It is the same dataset, but just a different split.
  2. The term “agent” in our documentation refers to other vehicles, not the ego vehicle.
  3. The data from annotation are human-annotated bounding boxes. So they can be considered ground-truth.