search

PavlosMelissinos / enet-keras

A keras implementation of ENet (abandoned for the foreseeable future)
MIT License
115 stars 46 forks source link

about the label format #20

Closed ShowTsai closed 4 years ago

ShowTsai commented 5 years ago

I have my own dataset There are 34774 png input images and 34774 png labels images input images are in shape (576, 576, 3) label images are in shape(576, 576) Every pixel in label image has a class number and there are 6 classes I don't quite understand how you deal with the MSCOCO annotations and what you do in "flow()" function in datasets.py, So what should I do on the"flow()"function , thanks!

PavlosMelissinos commented 5 years ago

Hey, apologies for taking so long to respond.

  1. I remember that the annotation system used by MS-COCO for semantic segmentation is kinda unusual if you're coming from a classification background. However, the gist is that datasets.py internally does some kind of conversion between alpha-like masks and polygons. Unfortunately you'll have to read the code in order to figure out exactly how it works, sorry about that :(
  2. flow is a function that lazily iterates over the dataset (it 'flows' over it) and is similar to keras.preprocessing.image.ImageDataGenerator.flow.

Other than that, is this issue still relevant to you or anyone else?

PavlosMelissinos commented 4 years ago

Other than that, is this issue still relevant to you or anyone else?

I suppose not... Closing it for the moment