Image area detection

[iharsuvorau]

(In this post, I'm summarizing what've been said in the chat on the topic)

Reasoning

Image area detection or picture-in-picture detection is all at the end the task of image segmentation. The difference with object detection is only what is highlighted by an algorithm. In our case we don't need any other objects except photographs itself. The end goal is us giving any picture, and if there's some other picture inside that picture then it should be highlighted. So, our task is also an object detection task. Basically, it's similar to something like this https://link.springer.com/article/10.1007/s11042-018-6387-5.

From the paper:

We provide trainable and testable datasetsFootnote 1 for art painting detection and identification. Although there are some datasets for object detection [9, 11, 19] of some generic objects like food, animals, vehicles, and datasets for art painting classification [10, 39] of genre and artist, as far as we know, there are no available datasets that can be used for art painting object detection and identification.

As expected, there's less data on picture-in-picture detection.

Approches

• FasterRCNN
• YOLO
• SSD (https://towardsdatascience.com/ssd-single-shot-detector-for-object-detection-using-multibox-1818603644ca)
• Detectron (https://research.fb.com/downloads/detectron/)

[iharsuvorau]

This tutorial on Detectron2, https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5#scrollTo=tjbUIhSxUdm_, and the part where they retrain the tool for ballons looks something that we can start with for picture-in-picture detection.