Is it convenient to provide more details and parameters of volume rendering?

[chenzyhust]

Such as near, far, n_samples, etc.

[GANWANSHUI]

Hi, we did not define the near and far like the bounded scene in NeRF, but we define the real size of the voxel. Here, we calculate the sampled point within the X ∈ (−52m, 52m), Z ∈ (0m, 6m), and Y ∈ (−52m, 52m).

self.Z = 16 self.Y = 256 self.X = 256 self.stepsize =0.5

N_samples = int(np.linalg.norm(np.array([self.Z // 2, self.Y // 2, self.X // 2]) + 1) / self.stepsize) + 1

The density of the sampled point out of the voxel would be set as 0, and the last sampled point is explicitly set as 1.

[chenzyhust]

Isn't voxel predicted in the lidar coordinate system? Why is the real size range of Z not [-5, 3] but [0, 6]

Hi, we did not define the near and far like the bounded scene in NeRF, but we define the real size of the voxel. Here, we calculate the sampled point within the X ∈ (−52m, 52m), Z ∈ (0m, 6m), and Y ∈ (−52m, 52m).

self.Z = 16 self.Y = 256 self.X = 256 self.stepsize =0.5

N_samples = int(np.linalg.norm(np.array([self.Z // 2, self.Y // 2, self.X // 2]) + 1) / self.stepsize) + 1

The density of the sampled point out of the voxel would be set as 0, and the last sampled point is explicitly set as 1.

Isn't voxel predicted in the lidar coordinate system? Why is the real size range of Z not [-5, 3] but [0, 6]

[GANWANSHUI]

In our experiment, we define the Z above the ego coordinate. You may also set the range in [-5, 3]. The voxel is defined in the ego coordinate, not the lidar coordinate.

[chenzyhust]

One more question，Is the rendering 95ms in the paper run on the existing cuda acceleration implementation or torch?

[GANWANSHUI]

just with the torch implementation