def detect(self, image):
"""
Find the object's new position in image, using current model M
"""
# Convert to grayscale
image = np.sum(image, 2) / 3
patch = crop_patch(image, self.region)
P = fft2(MOSSETrackerGrayscale.normalize(patch))
response = ifft2( np.conjugate(self.M) * P )
r, c = np.unravel_index(np.argmax(response), response.shape)
# Keep for visualisation
self.last_response = response
r_offset = np.mod(r + self.region_center[0], self.region.height) - self.region_center[0]
c_offset = np.mod(c + self.region_center[1], self.region.width) - self.region_center[1]
self.region.xpos += c_offset
self.region.ypos += r_offset
# Revert update if bbox is completely out of image
if self.region.intersection_box(BoundingBox("tl-size", 0, 0, image.shape[1], image.shape[0])).area() == 0.0:
self.region.ypos -= r_offset
self.region.xpos -= c_offset
return self.region
def test_crop_shape():
from cvl.image_io import crop_patch
from cvl.dataset import BoundingBox
import numpy as np
img = np.random.uniform(size=(1200, 600))
region = BoundingBox('tl-size', 50, 60, 100, 200)
crop = crop_patch(img, region)
assert crop.shape == region.shape()
def test_crop_pad_value_negative():
from cvl.image_io import crop_patch
from cvl.dataset import BoundingBox
import numpy as np
img = np.random.uniform(size=(1200, 600))
region = BoundingBox('tl-size', -1, 0, 100, 75)
crop = crop_patch(img, region)
assert crop[0, 0] == 0
def test_crop_region():
from cvl.image_io import crop_patch
from cvl.dataset import BoundingBox
import numpy as np
img = np.random.uniform(size=(1200, 600))
region = BoundingBox('tl-size', 60, 50, 100, 200)
crop = crop_patch(img, region)
assert crop[0, 0] == img[region.ypos, region.xpos]
assert crop[-1, -1] == img[region.ypos + region.height - 1, region.xpos + region.width - 1]
def update(self, image):
"""
Re-fit model M using new object position found in self.region (from detection step)
"""
# Convert to grayscale
image = np.sum(image, 2) / 3
patch = crop_patch(image, self.region)
normalized_patch = MOSSETrackerGrayscale.normalize(patch)
C = MOSSETrackerGrayscale.get_fourier_transformed_gaussian(height=self.region.height, width=self.region.width, std=self.std,
mean_x=0, mean_y=0)
P = fft2(normalized_patch)
self.A = self.A * (1-self.learning_rate) + np.conjugate(C) * P * self.learning_rate
self.B = self.B * (1-self.learning_rate) + np.conjugate(P) * P * self.learning_rate
self.M = self.A / self.B
return normalized_patch
def start(self, image, region):
"""
Construct initial model (=filter) in fourier domain using provided region in image
"""
# Convert to grayscale
image = np.sum(image, 2) / 3
# Where the gaussian should be centered
self.region = region
self.region_center = (region.height // 2, region.width // 2)
C = MOSSETrackerGrayscale.get_fourier_transformed_gaussian(height=region.height, width=region.width, std=self.std,
mean_x=0, mean_y=0)
patch = crop_patch(image, region)
P = fft2(MOSSETrackerGrayscale.normalize(patch))
self.A = np.conjugate(C) * P
self.B = np.conjugate(P) * P
self.M = self.A / self.B
def crop_patch(self, image):
region = self.region
return crop_patch(image, region)