Importance score for points

[charmeleonz]

Hi, thanks for the work. Just wondering what the values in Fig. 3(b) actually denote? (e.g. [0.3, 0.2 0.1, ...], from which 0.3 is taken)

[RunpeiDong]

Hi @charmeleonz,

Thanks for your interest.

• The vectors in Fig. 3(b) middle denote per-point features, e.g., [0.3, 0.2, 0.1, ...].
• The values in Fig. 3(b) right denote the value after max_pooling, which is used as the importance score, e.g., 0.3 is the max pooled value of the example feature vector.

Hope this could help you.

[charmeleonz]

So why can the max pooled value be used as an importance score?

[RunpeiDong]

@charmeleonz,

• Feature saliency as importance measurement is an effective and widely adopted practice in image-based KD methods or other areas, this inspires us to transfer this strategy;
• Intuitively, the feature with a higher value may indicate more significant attention to scene understanding for the network, which aligns with the sparse nature of point clouds;
• For 3D point clouds, max-pooling is an effective method to obtain permutation-invariant features, which is very important for geometry understanding.
• For a more detailed understanding, please refer to the paper.

[charmeleonz]

@RunpeiDong Thanks!