What kinds of ground-truth labels are required for MCG training?

[qjadud1994]

I currently survey various segment proposal methods.

However, I could not catch what kinds of ground-truth labels (e.g., segmentation map, object boundary) are required for MCG training because I'm not familiar with Matlab code :(

Also, some papers argue that MCG is an unsupervised method, is it right?

If anyone knows about this, please let me know.

[jponttuset]

Hi @qjadud1994,

There are two learnt parts in MCG:

• The low-level boundary detector, which is trained on BSDS500
• The Pareto front parameter search and the objectness ranking (a simple random forest), which are both trained on PASCAL VOC 2021 (train).

So in some sense it's "very little supervised" since it's "zero shot" on e.g. COCO, and the segmentation itself is not trained (heuristic on the hierarchy of regions given by the UCM)

[qjadud1994]

Thank you for your reply @jponttuset !

Then, in summary, "MCG requires object contour (boundary) labels for the training" is it right? What I want to know is what kind of supervision sources are required in MCG.

[jponttuset]

Then, in summary, "MCG requires object contour (boundary) labels for the training" is it right?

Correct, but not on the target domain, just once completely class-agnosticly and then it transfers zero-shot to the datasets.

What I want to know is what kind of supervision sources are required in MCG.

I don't get this.

[qjadud1994]

Thanks a lot @jponttuset !

Now, I understand about it. It will be of great help to me.