is tempered in competition, grow in actual combat
gathers more than a thousand data scientists sign up to participate in
HP Z series algorithm competition is in full swing
here, exchange and exchange, demonstrate technical strength
here, learn quickly, improve algorithm skills
here, here, unlimited creativity, unlock the Jizhi Mirror Realm
combined with the power of Z unlock the Jizhi Realm
"HP Z series algorithm competition" registration countdown
To help everyone achieve good results
, provide a guide to problem-solving algorithm
to clarify ideas, innovative algorithm
welcomes data scientists to challenge!
[ Introduction to the competition questions ]
This HP Z series algorithm competition mainly focuses on satellite / drone remote sensing scenarios. The official competition will provide scene data, and the contestants will detect road vehicles, complete " target detection with angle information locally" and debug the algorithm, and submit the results online.
[ Algorithm Introduction to ]
The problem-solving idea brought to you this time is to use the YOLOv3 target detection method to complete the positioning and classification of various rotating vehicle targets based on the fusion of DRBox algorithm ideas.
.1 Introduction to YOLOv3 Algorithm
With the integration of various ideas of the object detection method, the YOLO series detection algorithms have gone through the iterative process from YOLOv1 to YOLOv2, and now YOLOv3.
YOLOv1 is a single-stage object detection algorithm proposed by Joseph Redmon et al. in 2015. It divides the input image into S×S grids, each grid is responsible for predicting the target where the center point falls. Specifically, the entire network finally outputs (B×5+C) S×S feature maps, where B means that each network will predict B borders, each border includes 5 predicted values, namely x, y, w, h and confidence; C means the C target probability shared by B borders. Among the real-time detection algorithms at that time, YOLOv1's detection effect was the best. However, its biggest disadvantage is that its detection accuracy is significantly lower.
Considering that YOLOv1 has poor detection accuracy, Joseph Redmon et al. proposed the updated YOLOv2 detection algorithm in 2017. YOLOv2 focuses on solving the errors in recall and positioning accuracy. Compared with YOLOv1, the following improvements have been made: it improves the image resolution of classification network training; the idea of anchor box in Faster R-CNN was introduced; the design of network structure was improved; the output layer uses convolution layer to replace YOLO's full connection layer, etc., thereby greatly improving the recognition accuracy, positioning accuracy and detection speed.
Figure 1-1 Comparison of performance of YOLOv3 and other methods on COCO
YOLOv3 is a more accurate detection method proposed based on YOLOv2, absorbing the ideas of multi-layer detection, FPN feature pyramids, and ResNet better basic networks in SSD. Its performance on COCO is shown in Figure 1-1.
Compared with YOLOv2, YOLOv3 has improved from the following aspects:
1) Multi-label classification is used to replace the previous multi-cross entropy classification: each candidate box is classified in a multi-label manner, that is, binary cross entropy is used to learn the candidate box category. This classification method is very effective when the predicted target categories are complex and there are overlapping conditions;
2) uses multi-scale features for prediction: borrowing from the ideas of SSD and FPN, YOLOv3 predicts on three feature maps of different scales, namely scale1 (13×13), scale2 (26×26) and scale3 (52×52).At the same time, on scale2 and scale3, the features after upsampling of high-level features and fused with low-level features are used for prediction;
3) Set more and more effective prior boxes: YOLOv3 continues YOLOv2 uses cluster statistics to obtain the settings of the width and height of the prior boxes. Due to the increasing number of prediction layers, YOLOv3 sets up a total of 9 prior boxes and is evenly distributed on the prediction layer of each scale (that is, for the prediction layer of each scale, each feature point corresponds to 3 prior boxes, 13×13 corresponds to 3 larger, 52×52 corresponds to 3 smaller, and 26×26 corresponds to the remaining 3);
4) Use a more powerful feature extraction network: YOLOv3 combines the ideas of ResNet and NIN to propose a more efficient backbone called darknet-53. While retaining relatively low computational volume, Darknet-53 has a classification accuracy comparable to ResNet-101 and even ResNet-152.
After YOLOv3, a small number of detection methods based on this framework appeared, as follows:
1)GIoU (CVPR 2019): In most object detection algorithms at present, L2, L1 or Smooth L1 Loss is often used to train border adjustments, but IoU is used to determine whether the target is detected during evaluation. There is obviously inconsistency between these two methods, that is, the above metrics cannot be used to determine the quality of IoU. In response to this problem, GIoU proposed a more universal measure, directly using GIoU for loss calculation and gradient back-transmission, so as to make the target detection accuracy higher.
2) Gaussian YOLOv3 (ICCV 2019): Refer to KL Loss, Gaussian YOLOv3 uses a Gaussian model to predict borders based on YOLOv3. That is, for each prediction box, it changes from the previous (
) to (
), where
represents the coordinates of the border and represents the uncertainty of each coordinate. When finally obtaining the target confidence, the border includes the product of the target probability, the target category probability and the border determinism, that is,
. Here, uses the average of the above four variance .
details can be found at: https://github.com/pjreddie/darkne
.2 Introduction to the idea of rotating box
Figure 1-2 DRBox (based on SSD) algorithm diagram
DRBox uses directional boxes to overcome the difficulties of horizontal detection boxes in remote sensing images. The orientation box is a rectangle with an angle parameter that defines its orientation. It requires five parameters to return to its position, size and orientation. Compared to the horizontal detection box, the orientation box surrounds the outline of the target more closely, containing fewer background pixels than the horizontal box, so classification between the target and the background is easier.
details can be found in:
(DRBox)
https://github.com/liulei01/DRBo
(DRBoxv2)
https://github.com/ZongxuPan/DrBox-v2-tensorflow
.3 Recommended ideas
Traditional YOLOv3 uses a rotating rectangular box to detect targets, and related operations also basically maintain this feature, such as the setting of a prior box, the mechanism for matching a prior box during training, and the calculation of IoU during NMS. In the modeling of rotating rectangles, the biggest difference is that the additional rotation angle θ is introduced. Therefore, the above-mentioned related operating procedures must be adjusted accordingly.
.3.1 Prior box Setting
In the original YOLOv3, it mainly contains the following prior boxes. Here, the basic network is darknet-53 and the network input size is 416×416 for example.
1)scale1 (13×13): On this scale feature map, each point corresponds to 3 prior boxes of different sizes, namely [116, 90], [156, 198] and [373, 326], as shown in Figure 3-3.
Figure 3-3 scale1 (13×13) Traditional rectangular prior box example diagram
2) scale2 (26×26): On this scale feature map, each point also corresponds to 3 prior boxes of different sizes, namely [30, 61], [62, 45] and [59, 119], as shown in Figure 3-4.
Figure 3-4 scale2 (26×26) Traditional rectangular prior box example diagram
prior box zooms in 4 times showing
3) scale3 (52×52): On this scale feature map, each point also corresponds to 3 prior boxes of different sizes, namely [10, 13], [16, 20] and [33, 23], as shown in Figure 3-5.
Figure 3-5 scale3 (52×52) Traditional rectangular prior box example diagram
A prior box zooms in 8 times showing
While keeping the size of the prior box unchanged on each scale, we add a rotation angle to each prior box, and the rotation angle starts from 0° to 330° at a 30° interval. The prior boxes with rotation angles of each scale are shown in Figures 3-6, 3-7 and 3-8, respectively.
Figure 3-6 scale1 (13×13) Rotating rectangular prior box example diagram
Figure 3-7 scale2 (26×26) Rotating rectangular prior box example diagram
prior box zoom in 4 times display
Figure 3-8 scale3 (52×52) Rotating rectangular prior box example diagram
prior box zoom in 8 times display
.3.2 The prior box matches the target box
YOLOv3. In the matching process, the prior box of each scale is initialized as the background area, that is, the ground-truth of the Objectness is 0; then for each target box, find the prior box that best matches it, and make the ground-truth of the Objectness the ground-truth of the Objectness be 1. At the same time, the gradient back-passing is performed according to the classification of the prior box and the border regression bias value, and then iterate and optimize it continuously.
Under the traditional rectangular prior box, the IoU measure is generally used to obtain the best match for each target. IoU is defined as follows:
Based on traditional IoU, the matching metric between rotating rectangles ArIoU introduces the difference between angles while maintaining the position and size of the box, which is as follows:
where A represents the target box and the rotation angle is
; B represents the prior box and the rotation angle is
;
represents the prior box B about the center point so that it has the same rotation angle as the target box A. Therefore, ArIoU is equivalent to the basis of traditional IoU, with
added.
In the matching process on each scale prediction layer, for each rotating rectangular target, first obtain its corresponding grid; then find the rotation prior box that best matches it in the grid. If the prior box is exactly at the current scale, the prior box matches the current target box, and gradient back-passing is performed according to the classification of the prior box and the border regression bias value, and then iterates and optimizes continuously.
.3.3 A prior box matches the target box
Under the detection of the original traditional rectangular box, YOLOv3 will make predictions on three scales, assuming that the scale is represented as
. At the same time, for the points on each scale feature map, M prior boxes will be set. Each prior box will predict
values, namely the border offset (
), confidence Ch and the probability of belonging to each category.
Although the specific definition of the loss function is not mentioned in the technical report of YOLOv3, based on the understanding of the source code and its previous series of detection algorithms, the loss function contains three parts. One is to calculate the prediction error of the border offset using the European distance loss, the other is to define the prediction error of the reliability C using logistic regression and loss, and finally to define the prediction error of the multi-label P classification using logistic regression loss. The specific form is expressed as follows:
, where, after
is added to the rotation angle, the border offset predicted by the network becomes (
). Therefore, the loss of the network increases the error of the prediction angle. In addition, because the prior box introduces an angle dimension, the network predictions also multiplied. The new loss form is as follows:
where,
.3.4 NMS process
In NMS, the measurement between the predicted rotation boxes adopts the ArIoU form similar to that in the matching process. Since the same position generally contains only one target during the actual prediction process, the prediction box with an angle difference of 180° is also suppressed. Therefore, in the measurement of this process, the orientation of the head and tail of the target is generally ignored.Ignore the backward metric as follows:
After understanding the algorithm guide,
is eager to try
to show your algorithm creativity?
Come and participate in the HP Z series algorithm competition
and compete with data scientists!
[Task Time ]
Competition Schedule Registration and Model Submission
September 29-November 24
Results Review
November 25-December 7
Results announced
December 8
[Task Reward ]
Rich bonus waiting for you to get
.1 million yuan bonus pool
will be selected by the organizing committee to divide
st prize 1 30,000 yuan bonus
second prize 2 prizes 20,000 yuan bonus
third prize 3 prizes 10,000 yuan bonus
scan the code to register for the competition ↓
