literature reviewing for figure chart classification

VisSieve / main

https://vissieve.github.io/main/documentation/site

1 stars 0 forks source link

literature reviewing for figure chart classification #32

Open DevinBayly opened 4 weeks ago

DevinBayly commented 4 weeks ago

I'm using zotero, but I think https://ieeexplore.ieee.org/abstract/document/9729541?casa_token=cqOpWFrPXUcAAAAA:cTAZC5WGiU6_VpypUQjsprGsL1q20jN6NEkS_PHPeQm6cFKU--Cjqiu84VLx1LKkyx_1532yKQ is a great read since its quite recent and aims to provide a labeled database of figures.

we might be able to have a deep learning system trained on this and then apply it to the figures we have from princeton 2022

DevinBayly commented 3 weeks ago

ok, can count on what the suggestion of the vis images project is. They trained a bounding box detection system that can be pointed at a multi image figure and pull out regions and types. This is probably what we want to use https://github.com/VisImages/visimages-detection,

especially look at https://github.com/open-mmlab/mmdetection/blob/main/demo/demo.jpg and the result https://github.com/VisImages/visimages-detection/blob/main/result.jpg

DevinBayly commented 1 week ago

it turns out that mmdet can help us get bounding boxes as well as chart types when using the visimages detection. Then setup will be handled in a different card #document vis images detection

The main goal was to take the 30 element long "chart type" estimate variable and embed that in 2 dimensions.

the result at the moment doesn't immediately convey that the technique is working, but it does mean that we are getting closer.

I'll close with the comment that the visimages paper is definitely what we needed, but that Iqbal is recommending that we still use a different similarity method for exploring the images