Target Detection anchor-based and anchor free
1. anchor-free anchor-based differences and
Deep learning target detection usually is modeled in pairs some of the candidate regions of classification and regression problems. In a single stage detector, the candidate region is generated by the sliding window embodiment anchor; two-stage detector, the candidate region is RPN generated Proposal, but RPN itself is still produced on the sliding window embodiment of the anchor classification and regression .
anchor-free detection problem is solved by another means. Also divided into two sub-problems, i.e. to determine the center of the object and the four predicted frame. Center of the object is predicted, the predicted center of the target into which class prediction, can also be predicted centerness score of a soft. For predicting four border, it is more consistent, are forecast to ground truth the pixel frame from four sides, but will use some trick to limit the scope of regress.
2. anchor-free induction class algorithm
The key method based on multi-point joint expression of A.
a.CornerNet / CornerNet-lite: the upper left corner + lower right corner
b.ExtremeNet: up and down about four extremum points + center point
c.CenterNet: Keypoint Triplets for Object Detection: the upper left corner + lower right corner + center point
d.RepPoints: . 9 to adaptive learning samples beating
e.FoveaBox: center point + left corner + lower right corner
f.PLN: . 4 corner points + center point
B. prediction method based on a single central point
a.CenterNet: Objects as Points: center point + Width + Height
b.CSP: center point + height (the aspect ratio of a preset fixed target, based on the height calculated width)
c.FCOS: center point + of the frame 2 from
3. Why can anchor-free comeback
anchor-free approach can match anchor-based methods in accuracy, I think the greatest achievement should be attributed to FPN, secondly attributed Focal Loss. (Heart OS: high RetinaNet race). In the case of each location only predict a box, FPN structure of the scales played well compensated, FocalLoss is of great help to predict the central area. Of course, the method of transfer work is not so easy to do, I believe that some of the details will have considerable impact, such as handling overlapping areas, restrictions on the scope of the return, how will target assign to different FPN level, head whether share parameters, etc. .
4. anchor-free 和 single anchor
each anchor-free location mentioned above in the form of a square anchor may be equivalent, FCN is the use of a frame structure prediction (including the location and category) for each position of the feature map. However, anchor-free still makes sense, we can call anchor-prior-free. In addition, although both formal equivalence, but the actual operation is still a difference. In the anchor-based method, although each location may
Only one anchor, but the prediction is based on the object to match the anchor, and the anchor-free method is usually based on this matching point.
3. anchor-free of limitations
Although the accuracy of several methods above can be comparable with RetinaNet, but there is no obvious advantage (there may speed), and from two-stage cascade method still far difference. And single phase detector as the anchor-based, instance-level feature representation is not as good as the two phase detectors, the pattern in the head above will be relatively less. By the way Tucao about, above a small number of paper in order to achieve better-looking results, experimentally or hide some of the details there are not a fair comparison.
6. The object detection algorithm is generally divided into anchor-based, anchor-free, two converged, the difference is that there is no candidate object using the extracted anchor block.
A. anchor-based algorithm is based on behalf fasterRCNN, SSD, YoloV2 / V3 and the like.
fasterRCNN- provided three kinds of scale three kinds of the aspect ratio a total of nine anchor extraction candidate block
B. anchor-free algorithm is based on behalf CornerNet, ExtremeNet, CenterNet, the FCOS the like .
CornerNet- direct prediction is the probability that each point in the upper left, lower right corner of the goal box by pairing extract the upper left lower right corner
C. Fusion anchor-based and method anchor-free branch: FSAF, SFace, GA-RPN and so on.
FSAF- both a priori set anchor-based branches, there are anchor-free branch reinforcing abnormality detection capability target ratio
7. anchor (also referred to as anchor box) is prior to training, the training set using the method of k-means clustering and the like out of a set of rectangular frame represents the set of target length and width dimensions mainly. Characteristically FIG reasoning generated when these sliding anchor extracted n-th candidate rectangle do further classification and regression (refer to the detailed description presented anchor thought fasterRCNN article). That is the traditional target detection algorithm, the image pyramid use in the m * n in the sliding window. FIG conventional method wherein only different scale, generally fixed sliding window size; and similar fasterRCNN wherein FIG algorithm are fixed, Anchor multi-scale.
Three typical flowchart of object detection framework
8. Yolo a total of three versions of the algorithm, YoloV1 can be counted as anchor-free type; YoloV2 been introduced Anchor, YoloV3 also used anchor.