beam_id of my other datdsets

[wanghangege]

hello, thanks for your outstanding work, I have a doubt about make other datasets of their beam id labels , such as dair-v2x, it has 300 beam lidar; when i run this kitti_beam_id.py , it only classify four or five beams, the key code as follow:

``def get_beam_id(pc_index, path):
print(f'processing {pc_index}')
pc_kitti = np.fromfile(os.path.join(path, 'velodyne', "%06d.bin"%(pc_index)), dtype=np.float32).reshape(-1,4)
beam_id = np.zeros(pc_kitti.shape[0], dtype=np.float32)
selected = True
beam_count=0
cloud_size = pc_kitti.shape[0]
selectcount = 0
index = [0] * 65
for i in range(cloud_size):
beam_id[i] = beam_count
angle = np.arctan2(pc_kitti[i,1], -pc_kitti[i,0]) / np.pi * 180
if i==0:
last_point_angle = angle
back_point_angle = angle
if selected and selectcount<10:
selectcount += 1
else:
selected = False
if ((angle-last_point_angle)>=90) and (not selected) and ((angle-back_point_angle)>=90):
beam_count += 1
index[beam_count] = i
selected = True
selectcount = 0

    back_point_angle = last_point_angle
    last_point_angle = angle
    
if beam_count<63:
    for i in range(beam_count,64):
        index[i] = index[beam_count]
    print("scans less than 64")
elif beam_count>=64:
    print("gg, sort failed")
    return 
beam_id.tofile(os.path.join(path, "beam_labels", "%06d.bin"%(pc_index)))``

I have known kitti has 64 beams, and i changed the 64 to 300, is it right? could you help me...

[WoodwindHu]

Hello, different datasets have different organization mode of LiDAR beams. If you want to get the beam id of dair-v2x, you You'd better check its official documentation.

[wanghangege]

Hello, different datasets have different organization mode of LiDAR beams. If you want to get the beam id of dair-v2x, you You'd better check its official documentation.

Hello, thank you for your reply. As shown in the picture, I have three questions to consult you:

1. Does "oracle 84.8/71.6" in the figure refer to the average of difficult, general, and easy modes on Kit's test set dataset using Pointpillar? (BEV/3D);

2. The "ours 79.5/51.8" in the figure refers to the results obtained from training on Nuscences and testing on Kitti. However, when I used the model you published for testing, the actual results were much higher. May I ask if there was a mistake in my operation?
3. What does "closed gap" mean? I didn't see the explanation in the paper.

I would greatly appreciate it if you could reply to my question.

[WoodwindHu]

Does "oracle 84.8/71.6" in the figure refer to the average of difficult, general, and easy modes on Kit's test set dataset using Pointpillar? (BEV/3D);

It refer to the middle difficulty on kitti's val set under an IoU threshold of 0.7 over 40 recall positions.

May I ask if there was a mistake in my operation?

As in your figure, the performance of "ours" is "80.2/55.7". Public code is slightly different from experimental code for the sake of cleanliness, maybe this leads to a increase of performance.

What does "closed gap" mean?

Actually, we have explained it in the paper:

[wanghangege]

Thank you. The last one was indeed my negligence. I'm sorry, thank you again for answering my question