how to obtain the pkl

[tangtaogo]

Thank you for your great work. I would like to ask how you obtained the data_nuscene.pkl file. Specifically, how did you obtain the specific values for velocity, acceleration, and angular velocity? I noticed that the values I generated from the https://github.com/hustvl/VAD code repository are not consistent with the ones you provided. I suspect it may be because VAD generates values in the global coordinate system, while we are using the lidar coordinate system here. However, even after the conversion, they still don't match. I look forward to your response.

[buaazeus]

I have the same question.
How to get "fengze_nuscenes_infos_train.pkl", etc?
Thank you.

[FlyingSong888]

Have you solved this problem? I failed to reproduce the performance in ADMLP by using the original velocity and acceleration in canbus, and these message in canbus are different from that in data_nuscene.pkl file.

[FlyingSong888]

Have you solved this problem? I failed to reproduce the performance in ADMLP by using the original velocity and acceleration in canbus, and these message in canbus are different from that in data_nuscene.pkl file.

@tangtaogo @buaazeus @

[E2E-AD]

Hello, we would like to confirm the source of your code for obtaining velocity and acceleration. We have noticed some inconsistencies in previous projects, and we would like to use ST-P3 as the reference.

[FlyingSong888]

Hello, we would like to confirm the source of your code for obtaining velocity and acceleration. We have noticed some inconsistencies in previous projects, and we would like to use ST-P3 as the reference.

def _get_can_bus_info(nusc, nusc_can_bus, sample):
    scene_name = nusc.get('scene', sample['scene_token'])['name']
    sample_timestamp = sample['timestamp']
    try:
        pose_list = nusc_can_bus.get_messages(scene_name, 'pose')
    except:
        return np.zeros(18)  # server scenes do not have can bus information.
    can_bus = []
    # during each scene, the first timestamp of can_bus may be large than the first sample's timestamp
    last_pose = pose_list[0] # ['pos', 'orientation', 'accel', 'rotation_rate', 'vel']
    for i, pose in enumerate(pose_list):
        if pose['utime'] > sample_timestamp:
            break
        last_pose = pose
    _ = last_pose.pop('utime')  # useless
    pos = last_pose.pop('pos')
    rotation = last_pose.pop('orientation')

    can_bus.extend(pos)
    can_bus.extend(rotation)
    for key in last_pose.keys():
        can_bus.extend(pose[key])  # 16 elements
    can_bus.extend([0., 0.])
    return np.array(can_bus)

I used the velocity and acceleration obtained by the code above, the velocity and acceleration in canbus are like two scalars. I want to know how do you generate the velocity and acceleration, thank you!

[E2E-AD]

It looks like the code you've sent is taken from the implementation of VAD. In order to verify the consistency of the data, we compared the physical data items in VAD and ST-P3 and found inconsistencies in all the basic data (historical trajectory, future trajectory GT, speed, acceleration). This could be due to different preprocessing mechanisms between these projects. If you find this confusing, you could pose questions to their projects. Our project is based on ST-P3 and generates the above-mentioned historical physical information offline, and we adopt some feature enhancement methods to improve performance, such as performing exponential moving averages.

[buaazeus]

Can you please share the code to produce fengze_nuscenes_infos_train.pkl filter_token.pkl stp3_traj_gt.pkl stp3_occupancy.pkl?
Thank you

[scok30]

Ok, some of the pkl files were generated by different interns in our group. We will try our best to retrieve these codes. Please be patient, thank you.

[E2E-AD]

We have first uploaded the code generate_fengze.py used to generate fengze_nuscenes_infos_train.pkl and fengze_nuscenes_infos_val.pkl, located in pytorch/admlp/. We will try our best to continue uploading other files.