blownhither/ReplaceNet

2951I Final Project

ReplaceNet

Salient Object Replacement in 2D Images
with direct supervision from image synthesis

Datasets

• Super small dataset of SOD (Salient Objects Dataset).
  Descriptions: http://elderlab.yorku.ca/SOD/#download
  • http://elderlab.yorku.ca/SOD/SOD.zip
  • https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/segbench/BSDS300-images.tgz
  • decompress into SOD/ and BSDS300/

  import numpy as np
  from load_data import load_parsed_sod
  images, masks = load_parsed_sod()
  assert masks[0].dtype == np.bool

Training

• harmonize.py
  • Deep Image Harmonization

Using Synthesizer

from synthesize import Synthesizer
from load_data import load_parsed_sod
from matplotlib import pyplot as plt
s = Synthesizer()
images, mask = load_parsed_sod()
si = s.synthesize(image=images[223], mask=mask[223], reference_mask=mask[53])
plt.imshow(si)
plt.show()