torch-contour 1.1.7

Differentiable contour to image operations with PyTorch

torch_contour

Example of output of contour to mask and contour to distance map on a polygon in the form of a circle when varying the number of nodes

Download

$pip install torch_contour

Overview of the Toolbox

1. Pytorch layers for differentiable contour (polygon) to image operations.

• Contour to mask
• Contour to distance map
• Draw contour
• Smooth contour

2. Pytorch functions for contour feature extraction.

• Area
• Perimeter
• Curvature
• Hausdorff distance

3. Function for NumPy arrays only to remove loops inside contours and interpolate along the given contours.

Pytorch Layers

This library contains 3 pytorch non trainable layers for performing the differentiable operations of :

1. Contour to mask
2. Contour to distance map.
3. Draw contour.
4. Smooth contour

It can therefore be used to transform a polygon into a binary mask/distance map/ drawn contour in a completely differentiable way.
In particular, it can be used to transform the detection task into a segmentation task or do detection with any polygon.
The layers in 1, 2, 3 use the nice property of polygons such that for any point inside the sum of oriented angle is $\pm 2\pi$ and quickly converge towards 0 outside.
The three layers have no learnable weight.
All they do is to apply a function in a differentiable way.

Input (Float) (layer 1, 2, 3, 4):

A list of polygons of shape $B \times N \times K \times 2$ with:

• $B$ the batch size
• $N$ the number of polygons for each image
• $K$ the number of nodes for each polygon

Output (Float) (layer 1, 2, 3):

A mask/distance map/contour drawn of shape $B \times N \times H \times H$ with :

• $B$ the batch size
• $N$ the number of polygons for each image
• $H$ the Heigh of the distance map or mask

Output (Float) (layer 4):

Contours of shape $B \times N \times K \times 2$ with :

• $B$ the batch size
• $N$ the number of polygons for each image
• $K$ the number of nodes for each polygon

Important:

The polygon must have values between 0 and 1.

Example:

from torch_contour.torch_contour import Contour_to_distance_map, Contour_to_mask, Draw_contour, Smoothing
import torch
import matplotlib.pyplot as plt

polygons1 = torch.tensor([[[[0.1640, 0.5085],
         [0.1267, 0.4491],
         [0.1228, 0.3772],
         [0.1461, 0.3027],
         [0.1907, 0.2356],
         [0.2503, 0.1857],
         [0.3190, 0.1630],
         [0.3905, 0.1774],
         [0.4595, 0.2317],
         [0.5227, 0.3037],
         [0.5774, 0.3658],
         [0.6208, 0.3905],
         [0.6505, 0.3513],
         [0.6738, 0.2714],
         [0.7029, 0.2152],
         [0.7461, 0.2298],
         [0.8049, 0.2828],
         [0.8776, 0.3064],
         [0.9473, 0.2744],
         [0.9606, 0.2701],
         [0.9138, 0.3192],
         [0.8415, 0.3947],
         [0.7793, 0.4689],
         [0.7627, 0.5137],
         [0.8124, 0.5142],
         [0.8961, 0.5011],
         [0.9696, 0.5158],
         [1.0000, 0.5795],
         [0.9858, 0.6581],
         [0.9355, 0.7131],
         [0.9104, 0.7682],
         [0.9184, 0.8406],
         [0.8799, 0.8974],
         [0.8058, 0.9121],
         [0.7568, 0.8694],
         [0.7305, 0.7982],
         [0.6964, 0.7466],
         [0.6378, 0.7394],
         [0.5639, 0.7597],
         [0.4864, 0.7858],
         [0.4153, 0.7953],
         [0.3524, 0.7609],
         [0.3484, 0.7028],
         [0.3092, 0.7089],
         [0.2255, 0.7632],
         [0.1265, 0.8300],
         [0.0416, 0.8736],
         [0.0000, 0.8584],
         [0.0310, 0.7486],
         [0.1640, 0.5085]]]], dtype=torch.float32)  

width = 200

Mask = Contour_to_mask(width)
Draw = Draw_contour(width)
Dmap = Contour_to_distance_map(width)
smoother = Smoothing(sigma=1)

plt.imshow(Mask(polygons1).cpu().detach().numpy()[0,0])
plt.show()
plt.imshow(Draw(polygons1).cpu().detach().numpy()[0,0])
plt.show()
plt.imshow(Dmap(polygons1).cpu().detach().numpy()[0,0])
plt.show()

smoothed_polygons1_ = smoother(polygons1)

Pytorch functions

This library also contains batch torch operations for performing:

1. The area of a batch of polygons
2. The perimeter of a batch of polygons
3. The curvature of a batch of polygons
4. The haussdorf distance between 2 sets of polygons

from torch_contour.torch_contour import area, perimeter, hausdorf_distance, curvature
import torch


polygons2 = torch.tensor([[[[0.0460, 0.3955],
         [0.0000, 0.2690],
         [0.0179, 0.1957],
         [0.0789, 0.1496],
         [0.1622, 0.1049],
         [0.2495, 0.0566],
         [0.3287, 0.0543],
         [0.3925, 0.1280],
         [0.4451, 0.2231],
         [0.4928, 0.2692],
         [0.5436, 0.2215],
         [0.6133, 0.1419],
         [0.7077, 0.1118],
         [0.7603, 0.1569],
         [0.7405, 0.2511],
         [0.6742, 0.3440],
         [0.6042, 0.4099],
         [0.6036, 0.4780],
         [0.6693, 0.5520],
         [0.7396, 0.6100],
         [0.8190, 0.6502],
         [0.9172, 0.6815],
         [0.9818, 0.7310],
         [0.9605, 0.8186],
         [0.8830, 0.9023],
         [0.8048, 0.9205],
         [0.7506, 0.8514],
         [0.6597, 0.7975],
         [0.5866, 0.8195],
         [0.5988, 0.9145],
         [0.6419, 1.0000],
         [0.6529, 0.9978],
         [0.6253, 0.9186],
         [0.5714, 0.8027],
         [0.5035, 0.6905],
         [0.4340, 0.6223],
         [0.3713, 0.6260],
         [0.3116, 0.6854],
         [0.2478, 0.7748],
         [0.1732, 0.8687],
         [0.0892, 0.9420],
         [0.0353, 0.9737],
         [0.0452, 0.9514],
         [0.1028, 0.8855],
         [0.1831, 0.7907],
         [0.2610, 0.6817],
         [0.3113, 0.5730],
         [0.3090, 0.4793],
         [0.2289, 0.4153],
         [0.0460, 0.3955]]]], dtype = torch.float32)  


area_ = area(polygons2)
perimeter_ = perimeter(polygons2)
curvs = curvature(polygons2)
hausdorff_dists = hausdorff_distance(polygons1, polygons2)

NumPy remove loops and interpolate

cleaner = CleanContours()
cleaned_contours = cleaner.clean_contours(polygons2.cpu().detach().numpy())
cleaned_interpolated_contours = cleaner.clean_contours_and_interpolate(polygons2.cpu().detach().numpy())