Some common concepts in object detection transfer learning, IOU, NMS Understanding

1. Migration Learning

Migration learning is also known as supervised pre-training (supervised pre-training), which we often call migration learning. For example, you already have a bunch of picture data labeled for the age of the face, trained a CNN to recognize the age of the human face. Then when you meet a new project task is: Face sex recognition, then you can use the trained age to identify the CNN model, remove the last layer, and then the other network layer parameters are copied directly, continue to train. This is the so-called migration learning, the simple point is to take a task to train the parameters, get another task, as the initial parameter value of the neural network, so compared to you directly using random initialization method, the accuracy can be greatly improved.

There are many good training data, but there are few labeling data for object detection, how to train high-quality models with a small number of annotated data, for example, we first imagenet the image data set of advanced line network image classification training. This database has a large number of callout data.

2, IOU (orthogonal ratio)

Object detection needs to locate the object's bounding box, just like the picture above, we not only need to locate the vehicle's bounding box we also have to identify the object inside the bounding box is the vehicle. For the positioning accuracy of bounding box, there is a very important concept, because our algorithm can not be fully matched with manual data, so there is a positioning accuracy evaluation formula: IOU.
IOU represents the overlap of bounding box and ground truth, as shown in the following figure:

A coincident degree IOU of A and B of a rectangle is calculated as:

iou= (A∩B)/(A∪B)

1 2

Is the size of the overlapping area of a and b in the rectangular frame A, B, and the area ratio of the set:

iou=si/(Sa+sb-si)

1 2

3. NMS

NMS is also a non-maximal value suppression. In recent years, common object detection algorithms (including RCNN, Sppnet, FAST-RCNN, FASTER-RCNN, etc.) will eventually find many rectangles from a single image that may be objects, and then classify the probabilities for each rectangle as a category:

Just like the picture above, locating a vehicle, the algorithm finds a bunch of boxes, and we need to determine which rectangles are useless.
The so-called non-maximal value suppression: first assume that there are 6 rectangular boxes, according to the classification of classifier classification probability of sorting, from small to large respectively, the probability of belonging to the vehicle is a, B, C, D, E, F.

(1) Judging from the maximum probability rectangle F, the overlapping degree of a~e and F IOU is greater than the threshold value of a set;

(2) Assuming that the overlap of B, D and F exceeds the threshold, then discard B, D, and mark the first rectangle, F, which we have retained.

(3) from the rest of the rectangle box A, C, E, select the probability of the largest e, and then determine the degree of overlap between E and A, C, overlap is greater than a certain threshold, then throw away; and Mark E is the second rectangle we have left.

Just keep repeating and find all the rectangles that have been preserved.