Added new augmentations for Pattern Distortion. #242 #272
Conversation
…feature#242 pull changes
ss756
changed the title
|
|
||
| def __init__( | ||
| self, | ||
| width=500, |
There was a problem hiding this comment.
@ss756 Are those width, height and ws adjustable? If yes, it will be better to make them into a tuple so that the final value will be selected randomly in the code later. Same with self.perlin too. It can be initialized as random, so that if self.perlin is random, the code will randomly set it as True or False.
There was a problem hiding this comment.
Yes I will integrate this today
| super().__init__(p=p) | ||
| self.width = width | ||
| self.height = height | ||
| self.n_points = random.randint( |
There was a problem hiding this comment.
So we need a different approach here. The actual randomized parameter should be initialized during the __call__ so that the parameter value will be different from each call.
| result = image.copy() | ||
| h, w, _ = result.shape | ||
| delaunay_mesh = self.apply_augmentation() | ||
| # if self.perlin: |
There was a problem hiding this comment.
Do we still need this section (line 221 to line 224)? Else we can just remove it. If you still need it, please include some related information, maybe something like # Future development or something else equivalent.
| def __call__(self, image, layer=None, force=False): | ||
| if force or self.should_run(): | ||
| result = image.copy() | ||
| h, w, _ = result.shape |
There was a problem hiding this comment.
This line could be just h, w = result.shape[:2] because if the input is in grayscale, it will be facing error.
| super().__init__(p=p) | ||
| self.imgx = imgx # width of the image | ||
| self.imgy = imgy # hieght of the image | ||
| self.n_rotation = random.randint( |
There was a problem hiding this comment.
Similar with the issue i mentioned above, so this should be initialized in the __call__.
|
|
||
| def apply_augmentation(self): | ||
| # Applies the Augmentation to input data. | ||
| pattern_image = Image.new("RGB", (self.imgx, self.imgy)) |
There was a problem hiding this comment.
Is there any particular reason to use PIL instead of Numpy here? I think this overall section (line 78 - 106) can be just in Numpy?
| c = int(round(255 * z / self.n_rotation)) # color | ||
| pixels[kx, ky] = (c, c, c) # # RGB value | ||
| # pattern_image = ImageOps.autocontrast(pattern_image) # increasing the contrast of the image | ||
| pattern_image.save( |
There was a problem hiding this comment.
Actually why we need to save the image and load the image again here?
There was a problem hiding this comment.
Yeah I have written a comment in Voronoi.py
We save it as a 300 dpi image and load it
| (perlin_x, perlin_y), | ||
| ) | ||
| image = Image.fromarray(img_array) | ||
| image.save("images/Voronoi_example.png", "PNG", dpi=(300, 300)) |
There was a problem hiding this comment.
This should be same with quasicrystal, why we need to save and load it again?
There was a problem hiding this comment.
Hi @kwcckw I went through the link you sent. As per my understanding, they are also doing the same thing by saving the image as 300 dpi and reading it from memory. Let me know if I have misunderstood it so that I can make the changes in the new PR
| :param window_size: height and width of patch, default value is 200 | ||
| :type window_size: int | ||
| """ | ||
| h, w, _ = image.shape |
There was a problem hiding this comment.
This line should be replaced by h, w =image.shape[:2].
| cv2.BORDER_CONSTANT, | ||
| value=[255, 255, 255], | ||
| ) | ||
| h, w, _ = image.shape |
There was a problem hiding this comment.
This line should be replaced by h, w =image.shape[:2].
| image[ | ||
| initial_coords[1] : initial_coords[1] + hy, | ||
| initial_coords[0] : initial_coords[0] + wx, | ||
| :, |
There was a problem hiding this comment.
We can just remove the : here, else it will facing error if input is grayscale. same with line 131 below.
| def __call__(self, image, layer=None, force=False): | ||
| if force or self.should_run(): | ||
| result = image.copy() | ||
| h, w, _ = result.shape |
There was a problem hiding this comment.
This line should be replaced by h, w = result.shape[:2], and line 179 below can be just :
result = result[self.ws : h + self.ws, self.ws : w + self.ws]
We might facing error with the additional : if the input is in grayscale. So you will need to test the augmentation with grayscale image too.
resolves #242
This PR contains code for pattern-based distortion using 3 different techniques - Voronoi, Delaunay, and Quasicrystals. Perlin Noise is added to the Voronoi and Delaunay Tessellations patterns to create smoother and organic-looking images mentioned in #242.
Made changes to the docs and benchmark results in README.md