类别置信度预测:类别的目标执行度不都是1, 是根据proposal的box与gt的IOU进行设置的,记作
I
t
I_t
It,这样通过maxPooling就可以起到NMS的作用
I
=
m
i
n
(
1
,
m
a
x
(
0
,
2
×
I
o
U
t
−
0.5
)
)
I = min(1, max(0, 2\times IoU_t - 0.5))
I=min(1,max(0,2×IoUt−0.5))
L
s
c
o
r
e
=
−
I
t
l
o
g
(
I
^
t
)
−
(
1
−
I
t
)
l
o
g
(
1
−
I
^
t
)
L_{score} = -I_t log(\hat{I}_t) - (1 - I_t)log(1 - \hat{I}_t)
Lscore=−Itlog(I^t)−(1−It)log(1−I^t)
预测时候的置信度计算: inference 阶段,两个阶段的置信度几何均值进行综合;
Q
^
t
\hat{Q}_t
Q^t 最后预测的置信度,
I
^
t
\hat{I}_t
I^t是上面的描述的意思,
Y
^
t
\hat{Y}_t
Y^t是
m
a
x
0
<
=
k
<
=
K
Y
^
p
,
k
max_{0<=k<=K}\hat{Y}_{p,k}
max0<=k<=KY^p,k第一阶段预测的置信度。
Q
^
t
=
Y
^
t
∗
I
^
t
\hat{Q}_t = \sqrt{\hat{Y}_t * \hat{I}_t}
Q^t=Y^t∗I^t
3. 实验支撑
记录一些关键实验的结论分析,具有启发性的实验和结论
使用nuScenes数据集,有一个新的评测NDS:is a weighted average of mAP and others attributes metrics, including translation, scale, orientation, velocity, and other box attributes.
PKL评价指标: The PKL metric measures the influence of 3D object detection for down-streamed autonomous driving tasks based on the KL divergence of a planner’s route (using 3D detection) and the ground truth trajectory.