Returns the directory object detection history
Next: Depth articles - History of target detection (ii) elaborate R-CNN target detection
Directory Contents
Depth articles - Target Detection History (a) with respect to the classical target detection
Depth articles - History of target detection (ii) elaborate R-CNN target detection
Depth articles - Target Detection History (c) elaborate SPP-Net Target Detection
Depth articles - target detection history (four) elaborate Fast R-CNN from the target detection Faster R-CNN
Depth articles - Target Detection History (five) elaborate SSD target detection
Depth articles - Target Detection History (vi) elaborate YOLO-V3 target detection
Depth articles - target detection history (seven) elaborate YOLO-V3 object code detection of the Detailed
Depth articles - Target Detection History (eight) elaborate CornerNet-Lite Target Detection
Papers Address:
In this section, the detection of Classical target, the next section target detection elaborate R-CNN
A. Classical object detection
1. the first part: the training set configuration
(1) Negative sample
①. Selective search methods using the fusion region
②. Calculated coincidence between the bounding boxes for each region and the region labeled ground truth, if the coincidence degree between the region A and the ground truth of between 20% to 50%, and A and any other already generated a negative sample coincidence is not greater than 70%, then a is adopted as the negative samples.
(2) The positive sample
Ground truth is, those areas marked by hand as a positive sample.
2. Part II: extracting features of each positive / negative samples
HOG features + bag-of-words characteristics, while increasing auxiliary SIFT, two color SIFT, Extended opponent SIFT, RGB-SIFT features four such features to add dimensions to a staggering 360,000.
3. Part III: SVM classifier training.
4. Part IV: feedback False Positive (false positive cases)
These "False Positive" collected, SVM training just to get the right value as its initial weights, on the secondary SVM training, classification accuracy of SVM through secondary training in general there will be some improvement.
The testing process
(1) First method with selective search candidate area on the test image obtained
(2) then extracted feature vector of each area
(3). Bake trained SVM soft classification
(4). These regions are sorted according to the probability value
(5). The probability value is less than 0.5 area removed
(6) For those values of probability greater than 0.5, the calculation of the degree of overlap between the IOU and each region scores higher than its area, if the degree of overlap is greater than 30%, put the removed area (non-maximum value inhibition).
(7) The last remaining area as a target area.
Returns the directory object detection history
Next: Depth articles - History of target detection (ii) elaborate R-CNN target detection