why my training yolov2 network can't work?

I trained my data set according to the official documentation, detecting people in the image, only one category. The pre-training network also uses resnet50. The network structure is consistent with the example, but after many trainings, people are still not detected. During this time I tried to use the clustering algorithm to change the size or number of the achor box, the value of MaxEpochs, but still not working. Why is this?

I post my code and data here:

%% 利用resnet50预训练网络进行特征提取
load('gTruthPerson.mat')

% Define the image input size.
ImageWeight = 320;
ImageHeight = 320;
imageSize = [ImageWeight ImageHeight 3];

% Define the number of object classes to detect.
numClasses = 1;
anchorBoxes = [
     176  67 
     150  52  
     113  42
     271  99 
      99  34 
]; % .*ImageWeight/1280
anchorBoxes = round(anchorBoxes);

baseNetwork = resnet50();
% Specify the feature extraction layer.
featureLayer = 'activation_40_relu'; % mobilenetv2 的是'global_average_pooling2d_1';

% Create the YOLO v2 object detection network. 
lgraph = yolov2Layers(imageSize,numClasses,anchorBoxes,baseNetwork,featureLayer);
doTraining=0;
if doTraining
    
    % Configure the training options. 
    %  * Lower the learning rate to 1e-3 to stabilize training. 
    %  * Set CheckpointPath to save detector checkpoints to a temporary
    %    location. If training is interrupted due to a system failure or
    %    power outage, you can resume training from the saved checkpoint.
    options = trainingOptions('sgdm', ...
        'MiniBatchSize', 24, ....
        'InitialLearnRate',1e-3, ...
        'MaxEpochs',20,...
         'Verbose',1, ...
         'VerboseFrequency',1,...
        'CheckpointPath', './save', ...
        'ExecutionEnvironment','gpu',...
        'Shuffle','every-epoch');    
    
    % Train YOLO v2 detector.
%      pretrained = load('save/yolov2_checkpoint__45__2019_08_02__16_44_13.mat');
%     detector = pretrained.detector;
    [detector,info] = trainYOLOv2ObjectDetector(gTruthPerson,lgraph,options);
    save yoloDetector.mat detector info
else
    % Load pretrained detector for the example.
    pretrained = load('yoloDetector.mat');
    detector = pretrained.detector;
    info = pretrained.info;
end


%% test
for i = 1:height(gTruthPerson)
    I = imread(gTruthPerson.imageFilename{i});
%     I = imresize(I,imageSize(1:2));
    % Run the detector.
    [bboxes,scores] = detect(detector,I);
    
    % Annotate detections in the image.
    if ~isempty(bboxes)
        I = insertObjectAnnotation(I,'rectangle',bboxes,scores);
        imshow(I)
    end
end

all images and annotation files is here: google drive

Answer 1

[Song Decn]

I have same issue with the detector. any updates?

[Soukali Khiereddine]

Me to, I have similar problem

[enrico campo]

same problem

[cui,xingxing]

From this example, it can be seen that for users, it is very inflexible to watch the internal network. There is a problem there and it is difficult to debug. The official just gives a "yolov2Layers" or "yolov2OutputLayer", so that users cannot understand the design in principle. !!!

The official example is a very common and simple scenario, which cannot be consistent with the actual application!