Not All Points Are Equal: Learning Highly Efficient Point-based Detectors for 3D LiDAR Point Clouds

Last updated Jul 14, 2022 Edit Source

• Turns out F-FPS from 3DSSD is still not good enough in preserving foreground points.
  • Introduced Class-aware Sampling

• Superior sampling method compared to F-FPS and D-FPS
• Using vanilla cross-entropy loss $$L_{cls-aware}=-\sum^{C}_{c=1}(s_i\log(\hat{s_i})+(1-s_i)\log(1-\hat{s_i}))$$
• $C:$ Number of catergories
• $s_i:$ One hot labels
• $\hat{s_i}:$ predicted logits

• Give higher weight to points near instance centroid $$Mask_i=\sqrt[3]{\frac{\min(f^,b^)}{\max(f^,b^)}\times \frac{\min(l^,r^)}{\max(l^,r^)}\times\frac{\min(u^,d^)}{\max(u^,d^)}}$$
• $f^∗, b^∗, l^∗, r^∗, u^∗, d^∗$ represent the distance of a point to the 6 surfaces (front, back, left, right, up and down)
• This mask is used during training via the ctr-aware loss $L_{ctr-aware}$. At inference we simply keep top $k$ points with highest scores. $$L_{ctr-aware}=-\sum^{C}_{c=1}(Mask_i\cdot s_i\log(\hat{s_i})+(1-s_i)\log(1-\hat{s_i}))$$

• bruh $$L_{cent}=\frac{1}{|\mathcal{F}+|}\frac{1}{|\mathcal{S}+|}\sum_i\sum_j (|\Delta\hat{c_{ij}}-\Delta c_{ij}|+|\hat{c_{ij}}-\overline{c_i}|)\cdot \mathbb{1}S(p{ij})$$
• where $$\overline{c_i}=\frac{1}{|S_+|}\sum_j\hat{c_{ij}}\quad,\quad\mathbb{1}{S}(p{ij}):\mathcal{P}\to{0,1}$$
• $|\mathcal{F}_+|$ = the total number of GT boxes used to predict centroids
• $|\mathcal{S}_+|$ = number of points used to predict the instance center
• $\Delta\hat{c_{ij}}$ = offset predicted to instance center
• $\Delta{c_{ij}}$ = GT offset from a point $p_{ij}$ to center point
• $\mathbb{1}_S$ = indicator fn to determine whether point is used
  • implementation: expand GT box by some amount, then include all points inside the expanded box.

Examine the $|\hat{c_{ij}}-\overline{c_i}|$ term a bit more:

• Each GT box has one of these.
• For each relevant point (i.e. points inside the expanded GT box)
  • take their average, that becomes $\overline{c_i}$
• then while looping over each predicted centroid, the term is calculated.
• overall effect seems to promote all assignment points inside a GT box to predict the same centroid location