(Tensorflow backend)
Unofficial implementation of LEDnet in tensorflow keras
data_dir = "/home/anish/data_dir/" #Root training and Val data dir
train_image = "train_images"
val_images = "val_images"
train_masks = "train_masks"
val_masks = "val_masks"
#do not change below directory to avoid directory conflicts
#original input dimension in paper : 1024x512
model_param = {"image_size" : (512,512),
"train_batch_size" : 2,
"val_batch_size" : 2,
"augument":False
}
Todo
- Implement encoder and decoder block
- Update pre-trained weights on Cityscapes or Mapillary dataset
- Implement shuffle blocks as per paper
- Use Focal loss functions on unbalanced segmentations to test the performance.(Not so good at present)
- experiment different dilations apart from official parameters to evaluate segmentation.
Author : josh.anish1@gmail.com