Results on object detection datasets
s
Data
Image Size
FLOPs (in billion)
Params (in million)
mAP @ 0.50:0.95
map @ 0.50
mAP @ 0.75
Link
2.0
COCO
300x300
1.1
5.81
20.80
37.05
20.36
here
2.0
COCO
512x512
3.2
5.81
24.54
42.04
25.21
here
Below table compares the results on the PASCAL VOC 2007 dataset. It is important to NOTE that we do not use COCO data for below models.
s
Data
Image Size
FLOPs (in billion)
Params (in million)
mAP
Link
2.0
07+12
300x300
0.9
4.75
70.27
here
2.0
07+12
512x512
2.5
4.75
75.01
here
See below for class-wise mAP
512x512
300x300
aeroplane
0.7793
0.7465
bicycle
0.8199
0.8113
bird
0.7269
0.6517
boat
0.638
0.5972
bottle
0.4414
0.3506
bus
0.8376
0.78
car
0.8416
0.7865
cat
0.8533
0.829
chair
0.5915
0.5151
cow
0.802
0.7194
diningtable
0.7317
0.722
dog
0.819
0.7653
horse
0.8575
0.8389
motorbike
0.8179
0.7986
person
0.7803
0.7214
pottedplant
0.5353
0.4697
sheep
0.7553
0.6836
sofa
0.7884
0.7414
train
0.8449
0.8073
tvmonitor
0.7407
0.7166
mAP
0.7501
0.7026