Visualization

[Adam1904]

thanks for great work. Ihave some questions:

1- How can I visualize real points and virtual points in BEV (Bird's Eye View) both with and without an image, similar to Figure 1 and Figure 3c and d? Could you please explain how du the figures visualized?

2- Could you please explain how Figure 4 was visualized?

3-and one more question please, in the paper:

For instance segmentation, we use CenterNet2 [73] which adds a cascade RoI heads [3] on top of the first stage proposal network. The overall network runs at 40 FPS and achieves 43.3 instance segmentation mAP on the nuScenes image dataset [2].

What hyperparameters did you use to get this result (43.3 instance segmentation mAP)? How much is the learning rate? How many iterations? How many images per batch? how many epochs?

[tianweiy]

1. figure 1 uses open3d. and only virtual points inside bounding boxes are highlighted. the right side of figure 1 is zoom in (in the ui) and screenshots. for figure 3, a,b,c are illustration plots I drew in keynotes, d is also a zoom in + screenshot.

2. figure 4 is just matplotlib or something similar. We collapsed the z-xis of all lidar points and draw the boxes.

3. I don't remember the details now but I think this config is similar to what we use https://github.com/xingyizhou/CenterNet2/blob/master/configs/nuImages_CenterNet2_DLA_640_8x.yaml

[Adam1904]

Is there any code for that, please? Figure 1 uses open3d, and only virtual points inside bounding boxes are highlighted. I tried using this code https://github.com/tianweiy/CenterPoint/blob/master/tools/visual.py, but it still doesn't work. I executed the following command:

python ./tools/visual.py --path ./dataa/nuScenes/samples/LIDAR_TOP_VIRTUAL/n008-2018-05-21-11-06-59-0400__LIDAR_TOP__1526915243047392.pcd.bin.pkl.npy

And I modified main in visual.py:

if __name__ == '__main__':
    parser = argparse.ArgumentParser(description="LIDAR_TOP_VIRTUAL")
    parser.add_argument('--path', help='path to visualization file', type=str)
    args = parser.parse_args()

    data = np.load(args.path, allow_pickle=True).item()
    virtual_points = data['virtual_points']

    pcd = o3d.geometry.PointCloud()
    pcd.points = o3d.utility.Vector3dVector(virtual_points[:, :3])

    o3d.visualization.draw_geometries([pcd])

However, it hangs and the result is not displayed. And how can i use detections and scores to plot boxes? any help? thank you for ur reply