笔者最近阅览一些目标检测方法总结,看到一篇众多算法的表现对比表格,觉得很不错,与大家分享一下:

FPS(Speed) index is related to the hardware spec(e.g. CPU, GPU, RAM, etc), so it is hard to make an equal comparison. The solution is to measure the performance of all models on hardware with equivalent specifications, but it is very difficult and time consuming.

Detector VOC07 (mAP@IoU=0.5) VOC12 (mAP@IoU=0.5) COCO (mAP@IoU=0.5:0.95) Published In
R-CNN 58.5 - - CVPR'14
SPP-Net 59.2 - - ECCV'14
MR-CNN 78.2 (07+12) 73.9 (07+12) - ICCV'15
Fast R-CNN 70.0 (07+12) 68.4 (07++12) 19.7 ICCV'15
Faster R-CNN 73.2 (07+12) 70.4 (07++12) 21.9 NIPS'15
YOLO v1 66.4 (07+12) 57.9 (07++12) - CVPR'16
G-CNN 66.8 66.4 (07+12) - CVPR'16
AZNet 70.4 - 22.3 CVPR'16
ION 80.1 77.9 33.1 CVPR'16
HyperNet 76.3 (07+12) 71.4 (07++12) - CVPR'16
OHEM 78.9 (07+12) 76.3 (07++12) 22.4 CVPR'16
MPN - - 33.2 BMVC'16
SSD 76.8 (07+12) 74.9 (07++12) 31.2 ECCV'16
GBDNet 77.2 (07+12) - 27.0 ECCV'16
CPF 76.4 (07+12) 72.6 (07++12) - ECCV'16
R-FCN 79.5 (07+12) 77.6 (07++12) 29.9 NIPS'16
DeepID-Net 69.0 - - PAMI'16
NoC 71.6 (07+12) 68.8 (07+12) 27.2 TPAMI'16
DSSD 81.5 (07+12) 80.0 (07++12) 33.2 arXiv'17
TDM - - 37.3 CVPR'17
FPN - - 36.2 CVPR'17
YOLO v2 78.6 (07+12) 73.4 (07++12) - CVPR'17
RON 77.6 (07+12) 75.4 (07++12) 27.4 CVPR'17
DeNet 77.1 (07+12) 73.9 (07++12) 33.8 ICCV'17
CoupleNet 82.7 (07+12) 80.4 (07++12) 34.4 ICCV'17
RetinaNet - - 39.1 ICCV'17
DSOD 77.7 (07+12) 76.3 (07++12) - ICCV'17
SMN 70.0 - - ICCV'17
Light-Head R-CNN - - 41.5 arXiv'17
YOLO v3 - - 33.0 arXiv'18
SIN 76.0 (07+12) 73.1 (07++12) 23.2 CVPR'18
STDN 80.9 (07+12) - - CVPR'18
RefineDet 83.8 (07+12) 83.5 (07++12) 41.8 CVPR'18
SNIP - - 45.7 CVPR'18
Relation-Network - - 32.5 CVPR'18
Cascade R-CNN - - 42.8 CVPR'18
MLKP 80.6 (07+12) 77.2 (07++12) 28.6 CVPR'18
Fitness-NMS - - 41.8 CVPR'18
RFBNet 82.2 (07+12) - - ECCV'18
CornerNet - - 42.1 ECCV'18
PFPNet 84.1 (07+12) 83.7 (07++12) 39.4 ECCV'18
Pelee 70.9 (07+12) - - NIPS'18
HKRM 78.8 (07+12) - 37.8 NIPS'18
M2Det - - 44.2 AAAI'19
R-DAD 81.2 (07++12) 82.0 (07++12) 43.1 AAAI'19

从这个表格里面可以总结出两点:

  • 最新的算法比如M2Det, PFPNet, RefineDet碾压包括二阶段方法在内的方法;
  • 目标检测拼的不是召回率,而是特征提取的准确度,所以你看PFPNet这样的特征金字塔网络效果就非常的好。

从工程实践的角度来评价一下这些方法的实际使用感受:

  • 如果要说最好的方法,我觉得SNIP和PFPNet以及RFBNet不错;
  • M2Det和Pelee都比较难训练,但M2Det在小目标上的效果检测很强;
  • RefineDet各方面都很强,但感觉网络略微复杂

最后我们会在近期开源一个单阶段目标检测算法的仓库,我们将把M2Det,RFBNet,Pelee,RefineDet等最新优秀的方法集成到一个框架之中。