TensorFlow implementation of SegCaps [1]
I'm now testing this implementation with MSCOCO Dataset
- python 3.6 (though works both on python 3.6 and 2.7)
- Tensorflow >= 1.4
- numpy
- scipy
- skimage
- cv2
Before use, first check options in config.py
Options
batch_size : The number examples in a batch
max_iter : The number of iteration of training
mask : Whether to use mask in reconsturction loss.
mask_inv : Whether to apply color inversion to the image intensity.
- trainining:
python train.py --device 0 --mask True --mask_inv True
- deployment:
python deploy.py --device 0 --mask True --mask_inv True
- Custom Dataset in this case MSCOCO Dataset used
(The result seems too much overfitted, I'm now testing the several data augmentations.)
[1] Capsules for Object Segmentation (https://arxiv.org/abs/1804.04241)
Inwan Yoo / iwyoo@lunit.io modifications: Oushesh Haradhun oushesh@gmail.com