Target segmentation learning

object detection ： Pay more attention to semantic level
Target segmentation ： Focus on pixel level

Basic process of target detection algorithm

DPM： Additional strategies will be added , The pinnacle of traditional algorithms

Deep learning target detection method ：
One-stage(YOLO SSD)
Two-stage(Faster RCNN)

At present, deep learning target detection method is the mainstream
Acquire features through learning methods , Features will be more robust
Proposal Or go straight back to Get the target
When deciding the target , Through deep network
The feature is not robust ,
End to end , The end-to-end step is more convenient to adapt to the model than the multi-step of the traditional target detection algorithm

The basic flow of traditional target detection methods
Many classification ： Distinguish the background
After determining the candidate box

NMS： Not very much ： Merge candidate boxes

Deep learning objectives
Turn the process of feature extraction into convolutional neural network

Viola-Jones（ Face detection ）
haar features ：
Candidate box ： Sliding window extraction

HOG+SVM
For pedestrian detection （ There are differences in pedestrian posture , And for sports ）
extract HOG features
HOG features （
A kind of texture feature
Grayscale the features （Hog It is mainly used in gray image （
Conduct Gamma Transformation
Calculate the gradient map（ The direction angle of the current pixel ： Quantify to 0——360
dimension ： Quantitative angle , Degree of ,CELL,
Training SVM classifier
Modeling pedestrians and backgrounds
Support vector
Candidate target extraction ;
Candidate box filtering ：SVM

DPM： The pinnacle of traditional target detection algorithms
HOG An extension of
utilize SVM Train to get the gradient of the object

NMS：
Purpose ： To eliminate redundant frames , Find the best location for object detection
thought ： Select the window with the highest score in those fields , At the same time, suppress those windows with low scores
（ fraction ： Probability value ）
Soft-NMS： There are too many rough judgments of the detection frame by the direct standby threshold . therefore soft
Adjust the score of the detection maniac in the adjacent area, and it is not completely suppressed , Thus, the accuracy rate under the condition of high detection rate is improved .
It can still significantly improve the object detection performance at low retrieval rate .

ROI_POOLINT： Cutout
Two-stage：
RCNN
Fast RCNN
Faster RCNN
Faster RCNN variant