数式による物体形状の乱択化
: Initial point
: Transformed point
: Point movement
3D fractal model
Variance
check
Ground truth
generation
N
iteration
Alignment
3D bounding box & Centroid
3D fractal scene generation
3D IFS parameter setting & Affine transform
x
y
z
Intra-category augmentation
!!
=
$"
%"
&"
'"
("
)"
*"
ℎ"
,"
!!#$
+
."
/"
0"
!!
= −0.40, (!
= −0.61, +!
= 0.72,
/!
= −0.19, 1!
= −0.20, 2!
= −0.22,
3!
= 0.96, ℎ!
= −0.84, 6!
= −0.53,
9!
= −0.48, :!
= −0.79, ;!
= 0.83
1
( . = 1,2 … 1)
After M categories defined
Category 1
•••
Category M-2 Category M-1 Category M
Category 2 Category 3
Fractal category definition
Main: Category M
Noise: Category 2
Instance
augment
フラクタルにより3D形状をランダム⽣成しシーンに配置.
少数データをプリトレーニングに活⽤すると3D点群からの
物体検出(by VoteNet)性能が向上.
Ryosuke Yamada, et el., “Point Cloud Pre-training with Natural 3D Structure”, CVPR 2022