How to detect multiple stop sign using RCNN?

Hi,

I've tried the code given in the documentation as below.

Detect objects using R-CNN deep learning detector - MATLAB (mathworks.com)

It work. Then I try to change the testing image with this image (attachment). But the code only detect one stop sign as the result attachment.

I've tried to comment one line in the code that display only the stongest score. But it still detect one stop sign. supposedly there are 3 stop sign that can be found in the image.

% Display strongest detection result.
[score, idx] = max(score);

How can I improve the code to make it able to detect multiple stop sign?

Answer 1

[yanqi liu]

yes，sir，please use

[bboxes,scores] = detect(detector,I,'Threshold',0.1);

to get more rects，and then filter by score value

[Syukri Yazed]

Hi Yangqi Liu..

from your anwer which means, I should change this

% Display strongest detection result.
[score, idx] = max(score);

into this.

[bboxes,scores] = detect(detector,I,'Threshold',0.1);

Thus,

The whole code will be like this

load('RcnnNet1.mat', 'RcnnNet')
% Test the R-CNN detector on a test image.

img = imread('stopsign2_many.PNG');

% [bbox, score, label] = detect(RcnnNet, img, 'MiniBatchSize', 32);

% Display strongest detection result.
[bboxes, scores] = detect(RcnnNet, img, 'Threshold', 0.1);
% [score, idx] = max(scores);

% bbox = bbox(idx, :);
% annotation = sprintf('%s: (Confidence = %f)', label(idx), score);
annotation = sprintf('%s: (Confidence = %f)', scores);

detectedImg = insertObjectAnnotation(img, 'rectangle', bboxes, annotation);

figure
imshow(detectedImg)

I've try this, but it has an error.

Error using rcnnObjectDetector/parseDetectInputs (line 1225)
The argument 'Threshold' is a string scalar or character vector and does not match any parameter names. It failed validation for the
argument 'roi'.

Error in rcnnObjectDetector/detect (line 766)
            params = rcnnObjectDetector.parseDetectInputs(I, varargin{:});

Error in UseNLoadRcnnNetAgain (line 15)
[bboxes, scores, label] = detect(RcnnNet, img, 'Threshold', 0.1);           

[yanqi liu]

sir，if use rcnn

[bboxes, scores, label] = detect(RcnnNet, img, 'NumStrongestRegions', inf);

if use faster rcnn

[bboxes, scores, label] = detect(RcnnNet, img, 'Threshold', 0.1);

[Syukri Yazed]

I'm using RCNN, thus I'm using your first suggestion.

load('RcnnNet1.mat', 'RcnnNet')

img = imread('stopsign2_many.PNG');

[bboxes, scores, label] = detect(RcnnNet, img, 'NumStrongestRegions', inf);
[score, idx] = max(scores);
annotation = sprintf('%s: (Confidence = %f)', label(idx), scores);
detectedImg = insertObjectAnnotation(img, 'rectangle', bboxes, annotation);

figure
imshow(detectedImg)

the result is like this, (attachment 1 - result_NumStrongestRegion_inf.png). From the result it shows that all the detected regions have the same confidence level.

Then, I change 'inf' to '1000', just to make it to detect less. The result is as in the attachment (result_NumStrongestRegion_1000wB.png). From the result it shows there are three regions detected. But only 2 are correct. What I'm concern is the confidence scores are the same.

Thus, I think I need to add this line after [score, idx]

bboxes = bboxes(idx, :); % using detect(RcnnNet, img, 'NumStrongestRegions', 1000);

The result shows only one stopsign has been detected with score value = 1.

When using below (inf)

bboxes = bboxes(idx, :); % using detect(RcnnNet, img, 'NumStrongestRegions', inf);

The result shows wrong stopsign detection.

[yanqi liu]

yes，sir，it exist wrong detect result by rcnn，may be use faster rcnn or yolo to detect