深度学习: 从新视角 综述 Detection算法
Foreword
Continually updated，Constantly record my new summary of the Detection Algorithm。
Stage Handling
Stage Num | Algorithm |
|---|---|
4 | R-CNN、SPPNet、Cascade R-CNN |
2 | Fast R-CNN、Faster R-CNN、R-FCN、Mask R-CNN、Light-head R-CNN、SNIP、R-FCN-3000、SNIPER |
1 | YOLOv1、SSD、YOLOv2、DSSD、RetinaNet、YOLOv3 |
End-to-end Handling
End-to-end | Algorithm |
|---|---|
× | R-CNN、SPPNet |
√ | Fast R-CNN、Faster R-CNN、YOLOv1、SSD、R-FCN、YOLOv2、DSSD、Mask R-CNN、RetinaNet、Light-head R-CNN、SNIP、R-FCN-3000、YOLOv3、SNIPER |
RoI-wise subnetwork Handling
RoI-wise subnetwork | Algorithm |
|---|---|
× | YOLOv1、SSD、YOLOv2、DSSD、RetinaNet、YOLOv3、R-FCN、Light-head R-CNN、R-FCN-3000 |
√ | R-CNN、SPPNet、Fast R-CNN、Faster R-CNN、Mask R-CNN、SNIP、Cascade R-CNN、SNIPER |
Scale Handling
Scale Handling | Algorithm | Sample Diagram |
|---|---|---|
Image Pyramid | SNIP、SNIPER | |
Single Feature | R-CNN、SPPNet、Fast R-CNN、Faster R-CNN、YOLOv1、R-FCN、Light-head R-CNN、R-FCN-3000 | |
Pyramidal Feature | SSD、YOLOv2 | |
Feature Pyramid Network | FPN、DSSD、Mask R-CNN、Cascade R-CNN、PAN |
Repeatability Handling
Decrease repetition | Annotation |
|---|---|
SPPNet | 去除卷积的重复抽取 |
Fast R-CNN | 去除卷积的重复抽取 |
R-FCN | 去除RoI-wise network的重复计算 |
Light-head R-CNN | 去除每个类都要单独进行的location regression |
R-FCN-3000 | 去除每个类都要单独进行的location regression |
Increase repetition | Annotation |
|---|---|
SDD | 增加feature map的选取数量 |
YOLOv2 | 增加feature map的选取数量 |
FPN | 增加feature map的选取数量 |
SNIP | 增加image scale的数量 |
Cascade R-CNN | 增加RoI-wise network的数量 |
PAN | 增加feature map的选取数量 |
Basemodel Handling
Algorithm | AlexNet | ZFNet | GoogleNet | VGGNet | DarkNet | ResNet | FPN(+ResNet) | PAN(+ResNet) | DetNet |
|---|---|---|---|---|---|---|---|---|---|
R-CNN | √ | ||||||||
SPPNet | √ | ||||||||
Fast R-CNN | √ | ||||||||
Faster R-CNN | √ | √ | |||||||
YOLOv1 | √ | ||||||||
SSD | √ | ||||||||
R-FCN | √ | ||||||||
YOLOv2 | √ | ||||||||
FPN | √ | ||||||||
DSSD | √ | ||||||||
Mask R-CNN | √ | ||||||||
DCN | √ | ||||||||
RetinaNet | √ | ||||||||
Light-head R-CNN | √ | ||||||||
Relation Network | |||||||||
SNIP | √ | ||||||||
Cascade R-CNN | √ | ||||||||
R-FCN-3000 | √ | ||||||||
PAN | √ | ||||||||
YOLOv3 | √ | ||||||||
DetNet | √ | ||||||||
SNIPER | √ |
Classification Handling
Algorithm | SVM | Softmax loss | Focal loss |
|---|---|---|---|
R-CNN | √ | ||
SPPNet | √ | ||
Fast R-CNN | √ | ||
Faster R-CNN | √ | ||
YOLOv1 | √ | ||
SSD | √ | ||
R-FCN | √ | ||
YOLOv2 | √ | ||
FPN | √ | ||
DSSD | √ | ||
Mask R-CNN | √ | ||
DCN | √ | ||
RetinaNet | √ | ||
Light-head R-CNN | |||
Relation Network | |||
SNIP | |||
Cascade R-CNN | |||
R-FCN-3000 | |||
PAN | |||
YOLOv3 | |||
DetNet | |||
SNIPER |
Reg_loss Handling
Algorithm | L2 loss | Smooth L1 loss |
|---|---|---|
R-CNN | √ | |
SPPNet | √ | |
Fast R-CNN | √ | |
Faster R-CNN | √ | |
YOLOv1 | ||
SSD | √ | |
R-FCN | ||
YOLOv2 | ||
FPN | √ | |
DSSD | ||
Mask R-CNN | ||
DCN | ||
RetinaNet | √ | |
Light-head R-CNN | ||
Relation Network | ||
SNIP | ||
Cascade R-CNN | ||
R-FCN-3000 | ||
PAN | ||
YOLOv3 | ||
DetNet | ||
SNIPER |
Encode/Decode Handling
Role: normalize the target, making the regression task simple.
Algorithm | The earliest target | [xcwa,ycha,logw,logh][xcwa,ycha,logw,logh][\frac{x_c}{w_a}, \frac{y_c}{h_a}, \log{w}, \log{h}] | Others |
|---|---|---|---|
R-CNN | √ | ||
SPPNet | √ | ||
Fast R-CNN | √ | ||
Faster R-CNN | √ | ||
YOLOv1 | |||
SSD | √ | ||
R-FCN | √ | ||
YOLOv2 | √ | ||
FPN | √ | ||
DSSD | |||
Mask R-CNN | √ | ||
DCN | √ | ||
RetinaNet | √ | ||
Light-head R-CNN | √ | ||
Relation Network | √ | ||
SNIP | √ | ||
Cascade R-CNN | √ | ||
R-FCN-3000 | √ | ||
PAN | √ | ||
YOLOv3 | √ | ||
DetNet | √ | ||
SNIPER | √ |
腾讯云开发者
