nuScenes val split results file generation

[Kay1794]

Hi @TsuTikgiau , sorry for my delay in responding to this, I hope it's not too late to help! Here are the steps to get the challenge datasets in our data format:

1. Change the Scene constructor to the following:

    def __init__(self, timesteps, sample_tokens, map=None, dt=1, name="", frequency_multiplier=1,
                 aug_func=None, non_aug_scene=None, x_min=None, y_min=None):
        self.map = map
        self.timesteps = timesteps
        self.dt = dt
        self.name = name
        self.sample_tokens = sample_tokens  # These are ordered so that they correspond to each timestep in the scene.
        self.x_min = x_min  # Used to get predictions back into global coordinates.
        self.y_min = y_min  # Used to get predictions back into global coordinates.

        self.nodes = []

        self.robot = None

        self.temporal_scene_graph = None

        self.frequency_multiplier = frequency_multiplier

        self.description = ""

        self.aug_func = aug_func
        self.non_aug_scene = non_aug_scene

1. Re-run process_data.py with this updated version. This process_data.py is the exact file I used for the nuScenes prediction challenge, so if you’re using those data splits (their train, train_val, val splits) then this should run just fine.

I hope this helps!

Hi @BorisIvanovic, Thank you so much for providing the data generation code for the nuscenes challenge. I met a problem when I was about to generate result json file for challenge submission. I noticed that the number of output lines in the results csv file is 5000+ however the challenge requires 9041 lines of result. I followed your instruction from other two issues #13 and #11 for training/validation but the problem remained. Any idea about that?

Thank you!

Originally posted by @Kay1794 in #21 (comment)