How to label ground truth automatically for images

Hello,

I have 2500 images and would like to label ground truth automatically without manually drawing bounding boxes. Could anyone help to show me the codes and steps please?

[Cecilia Tan]

Hi Dahe, Oh great. Thanks for helping out Dahe!

Answer 1

[Ridwan Alam]

Automated data labeling is still an active research field.

Not sure if this would be helpful in the meantime:

https://www.mathworks.com/help/vision/image-labeling.html?s_tid=srchbrcm

[Cecilia Tan]

Hi Ridwan,

Thank you for your reply. Yes that's right, I did try to use Image Labeler but that requires me to manually draw bounding box around an object for each image. This will take a long time so I'm looking at Automated Algorithm but I don't know how. Could you provide some codes?

[Ridwan Alam]

I totally understand your frustration. We all had been there. But as Walter mentioned below, if we could automate the process of labelling, why would we need those labels for?

As a starter, you can manually label a few images > train a simple model > use that trained model to label some from the rest of the data > manually evaluate the performance, i.e. whether the model is doing a reasonable job in labelling > retrain with new data > reevaluate ..

As I said before, it's still an area of active research. Good luck!

Answer 2

[Walter Roberson]

Ground truth data is, by definition, as accurate as an expert in the field can possibly be. For example for an overhead picture of land, ground truth does not just assume that a green blob is a tree: ground truth would be a botonist (or forester) going through the area in person or in high resolution photos and verifying that the blob is not just green paint and not just just a dark plant, but is indeed a tree (and not, for example, a herbacious growth such as a banana, as banana are herbs, not trees (no wood)).

If there were an algorithm that could give a ground-truth labeling from whatever inputs you have available, then you would probably not need to do whatever it is you intend to do with the ground truth labels. Not unless you were working an an algorithm that was perhaps less perfect but was substantially faster.

[Cecilia Tan]

Hi Walter,

Thank you for your reply. I'm still new to Matlab. I think my problem is label ROI for each image so I can export the ground truth data of the object in the image. Is there a quicker way to label thousands of images?

[Walter Roberson]

Sometimes it can help to do edge detection to create an interactive roi to throw up on the graph for the user to fine tune.

However context is important. If the question is picking out where a ball is in a well lit box then things might be fine. If the question is picking out brain tumors on MRI then that is not something that you can ground truth from MRI images alone: ground truth for that involves microscopy and possibly oxygen flow studies and possibly MRS. Ground truth for brain tumor images is not whether the shade of gray of the pixel hints that there might be a tumor at that location: ground truth for brian tumor images is whether or not there is a brain tumor at that location, which medical images only approximate.