![](https://csdnimg.cn/release/download_crawler_static/12876921/bg4.jpg)
Table 2. Commonly used metrics for different tasks. In this table, N denotes the number of samples, C denotes the number of categories,
IDS denotes the number of identity switches, I
i,j
denotes the number of points that are from ground truth class/instance i and labelled
as j, T P/T N/F P/F N stands for the number of true positives, true negatives, false positives and false negatives respectively. Higher
metrics indicate better results if not specified otherwise.
Metric Formula Explanation
Accuracy Accuracy =
T P +T N
T P +T N+F P +F N
Accuracy indicates how many predictions are correct over all
predictions. “Overall accuracy (OA)” indicates the accuracy on
the entire dataset.
mACC mACC =
1
C
P
C
c=1
Accuracy
c
The mean of accuracy on different categories, useful when the
categories are imbalanced.
Precision P recision =
T P
T P +F P
The ratio of correct predictions over all predictions.
Recall Recall =
T P
T P +F N
The ratio of correct predictions over positive samples in the
ground truth.
F1-Score F
1
= 2 ×
P recision·Recall
P recision+Recall
The harmonic mean of precision and recall.
IoU IoU
i
=
I
i,i
P
C
c=1
(I
i,c
+I
c,i
)−I
i,i
Intersection over Union (of class/instance i). The intersection
and union are calculated between the prediction and the ground
truth.
mIoU mIoU =
1
C
P
C
c=1
IoU
i
The mean of IoU on all classes/instances.
MOTA MOT A = 1 −
F N+F P +IDS
T P +F N
Multi-object tracking accuracy (MOTA) synthesizes 3 error
sources: false positives, missed targets and identity switches,
and the number of ground truth (as T P + F N ) is used for nor-
malization.
MOTP MOT P =
P
i,t
e
i,t
P
t
d
t
Multi-object tracking precision (MOTP) indicates the precision
of localization. d
t
denotes the number of matches at time t, and
e
i,t
denotes the error of the i-th pair at time t.
EPE EP E = ||
ˆ
sf − sf||
2
End point error (EPE) is used in scene flow estimation, also
referred as EPE2D/EPE3D for 2D/3D data respectively.
ˆ
sf
denotes the predicted scene flow vector while sf denotes the
ground truth.
Figure 2. 3D models from ShapeNet [5]. ShapeNet contains large-scale 3D models with manually verified annotation.
tion of the grid is strictly limited. OctNet [84] improved the
efficiency by introducing a hybrid grid-octree structure to
hierarchically partition point clouds. A point cloud is repre-
sented by several octrees along a regular grid, each octree is
encoded as a bit string, and features are generated through
naive arithmetic. Inspired by OctNet, OCNN [99] then pro-
posed a method that introduces 3D-CNNs to extract features
from octrees.
4