[Note] Artificial Intelligence Section VI: evolutionary history YOLO target detection algorithm

YOLO works is that the picture input to the multi-layered convolution to extract picture feature. Then return to the target directly in the output layer frame coordinate its category belongs. Finally, by removing the overlapping NMS process target block.
Faster-RCNN with different, Faster-RCNN after extraction is a characteristic graph used to scan the candidate block diagram embodiment to find the target feature. And then return categorized by two sub-networks. A candidate block for adding an offset scaled to return to the target coordinates. A target used for classification. Finally, by removing the overlapping NMS process target block.

advantage:

Disadvantages:

Since the output directly from FIG classification characteristic, features, FIG lower resolution, corresponding to a high overlap, or small dense target object, the detection rate is low.

Network architecture:

Network input image size (448 * 448 * 3), and the maximum value convolutional pooled, to give (7 * 7 * 1024) feature.
Wherein the connecting layer through the full 4096, is converted into (7 * 7 * 30) results.
(7 * 7 * 30) results in the corresponding 7 * 7 = 49 feature points, can detect up to 49 targets in the picture. 30 here comprises two pieces of information, a portion of the target box is the two coordinates (x, y, w, h) and the confidence level classification part 20, and therefore (1 + 4) * 2 + 20 = 30.

Disadvantages:

Less intensive target low detection rate.
Target box is not accurate, there will be some deviation.
Large differences in the aspect ratio of the object, poor generalization capabilities.

Improvements:

Regression directly by the target coordinates (x, y, w, h), the candidate block size is set in advance to each of feature points corresponding to each of FIG. 3 in size and shape different candidate block, then return to the candidate frame offset . When such a large target detection can be optimized aspect ratio difference, increase the precision of the goal box, speed up training.
Image feature extraction network instead Darknet-19 network, providing training speed.
By the input image size (448 * 448 * 3) was changed to (416 * 416 * 3).
Extracted (13 * 13) feature FIG. Further taking the foregoing (26 * 26) wherein FIG converted into (13 * 13) a characteristic graph for predicting smaller objects.

Network architecture:

Docknet19, and finally the classification, a total of 1000 category. YOLO network feature extraction is removed as the end of the network:

And then (300 * 13 * 13) the convolution of the target frame coordinates and classification. (1 + 4 + 20) * 3 = 300.

Speed comparison:

Improvements:

The image processing speed may reach 608x608 20FPS on Pascal Titan X, in the [email protected] COCO test-dev reached 57.9%.
There are three kinds of input picture size (320 * 320), (416 * 416), (608 * 608). Rate from fast to slow, from low to high accuracy.
Use FPN pyramid feature extraction different scales of 3 layers, each candidate feature using 3 different size and shape of the frame.
Feature extraction network of Darknet-19 to Darknet-53, increasing the depth of the network.
Each block corresponding to the respective classification do Softmax, Logistic to a plurality of classifiers.
Category losses using binary cross-entropy loss.

Network architecture:

Darknet-53, and finally the classification, a total of 1000 category. YOLO network feature extraction is removed as the end of the network:

darknet-53 close to ResNet-101 or accuracy ResNet-152, but faster, the following comparison:

Speed comparison:

in conclusion:

YOLOv3 with good results on [email protected] and small target APs, but with the increase, performance degradation IOU, indicating YOLOv3 not fit well with the ground truth.

References:

PPHT H

Published 28 original articles · won praise 2 · views 10000 +