DAVAR-Lab-OCR icon indicating copy to clipboard operation
DAVAR-Lab-OCR copied to clipboard

Published 20 hours ago • verified publisher icon hikopensource

Need tips on how to visualize GPMA masks

Open xiangn95 opened this issue 2 years ago • 3 comments

Hi there! I need some help in visualizing the GPMA masks. When I visualized the LPMA masks, it look perfectly normal as shown below.

LPMA Horizontal Mask lpma_hor

LPMA Vertical Mask lpma_ver

But the visualization of GPMA masks are incomprehensible.

GPMA Horizontal Mask gpma_hor

GPMA Vertical Mask gpma_ver

The way I visualize GPMA mask is by first retrieving the global mask from result[3] of simple_test.py (as result[3][1] and result[3][2] are horizontal and vertical masks respectively), then I multiply the masks by 255. I tried multiplying the masks with the cell box (result[3][0]), but the result is still messy.

Hope to get some tips on how to improve the visualization of the GPMA masks. Thanks!

xiangn95 avatar Jun 09 '22 07:06 xiangn95

It looks like the model is not well learned, can you check if the prediction target of the model is correct (L163 in pipelines/gpma_data.py)?

qiaoliang6 avatar Jun 16 '22 12:06 qiaoliang6

The model is pretrained on pubtabnet provided by the authors, and the picture visualized is also a test image from pubtabnet. I tried visualize the training image as well but get the same result.

xiangn95 avatar Jun 16 '22 14:06 xiangn95

Hi! Did you manage to solve this problem? I am having the same issue. The global segmentation does't seem to work properly. I understand that the task mainly exists in order to improve the local tasks however if the global segmentation doesn't work I expect a decrease in performance even in the local segmentations and pyramids

LucaDeGrandis avatar Feb 28 '23 09:02 LucaDeGrandis