tensorflow - DeepLab Semantic Segmentation, replacing masks by coordinates

Question

I have some images which are segmented in several classes.

Instead of having complex zones, all of them are rectangles. So my dataset consists of:

[
   image,
   GT:[
      {
         x0:
         y0:
         x1:
         y1:
         label:
      }
   ]

I would like to use a pre-trained (on VOC_2012) Deeplab model and train it on my data. Knowing that the shapes are simple rectangles, am I still obliged to create a png mask for each image, or can I keep the data as is, and then create the mask on the fly ?

Would it be relevant/worth it ? (the dataset is consisting in 500K images, approximating 50GB in total)

Answer 1

等待大神答复