Multiple Instances of Same Object in Image - Object Detection Using CNN

Question

New to NN's. A CNN can be trained to detect a single object in an image. However, what if any image in a dataset could contain any n # of objects. Does this not pose a problem to CNNs as the output dense layer has to be a fixed size? How would you solve this problem?

For example: Let's say I randomly sampled 2 images from this set. Image 1 has 2 objects and image 2 has 5 objects. The y label for img1 would contain the bounding box coordinates for 2 objects; the y label for img2 would contain coordinates for 5 objects -- much larger y vector than img1.

A possible solution? :

I would need to find the image with the largest # of objects (designate this value as M). Let's also say an object has 4 coordinates. If M = 5, I would need a y vector of 20. If an image has 1 object, the y vector would contain 4 non-zero values AND 16 zero values. The 4 non-zero values would represent the coordinates and the 16 zero values would represent the coordinates of the other non-existent objects.

Reda El Hail · Accepted Answer · 2021-11-21 05:58:06Z

1

The basic way of doing multiple object classification is using segmentation. This is done by segmenting the input image to several sub-areas and feed each area to the neural network.

However, this is a very basic method and there are now many advanced algorithms that do segmentation automatically.

Generally, multiple object classification is tackled in two steps: First a region proposal algorithm to guess which parts of the image contains the object.

The second is an algorithm to classify the proposed regions.

img source

answered Nov 21, 2021 at 5:58

Reda El Hail

1,0161 gold badge9 silver badges18 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

Ayma Over a year ago

Something like the Faster R-CNN performs segmentation by using RPN to extract features it deems relevant? Is my understanding of this correct? Furthermore, what would the y vector look like given that there are variable prediction labels for images.

Collectives™ on Stack Overflow

Multiple Instances of Same Object in Image - Object Detection Using CNN

1 Answer 1

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related