def update(self, image):
"""
Given current image, returns target box.
"""
image = np.array(image)
if image.ndim == 2:
image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
# crop current and previous image at previous box location
prev_sample, opts_prev = crop_sample({
'image': self.prev_img,
'bb': self.prev_box
})
curr_sample, opts_curr = crop_sample({
'image': image,
'bb': self.prev_box
})
self.opts = opts_curr
self.curr_img = image
curr_img = self.scale(curr_sample, opts_curr)['image']
prev_img = self.scale(prev_sample, opts_prev)['image']
sample = {'previmg': prev_img, 'currimg': curr_img}
sample = self.transform_tensor(sample)
# do forward pass to get box
box = np.array(self._get_rect(sample))
# update previous box and image
self.prev_img = image
self.prev_box = np.copy(box)
# convert [xmin, ymin, xmax, ymax] box to [xmin, ymin, width, height]
# for correct evaluation by got10k toolkit
box[2] = box[2] - box[0]
box[3] = box[3] - box[1]
return box
def get_sample(self, idx):
"""
Returns sample without transformation for visualization.
Sample consists of resized previous and current frame with target
which is passed to the network. Bounding box values are normalized
between 0 and 1 with respect to the target frame and then scaled by
factor of 10.
"""
opts_curr = {}
curr_sample = {}
curr_img = self.get_orig_sample(idx, 1)['image']
currbb = self.get_orig_sample(idx, 1)['bb']
prevbb = self.get_orig_sample(idx, 0)['bb']
bbox_curr_shift = BoundingBox(prevbb[0], prevbb[1], prevbb[2],
prevbb[3])
(rand_search_region, rand_search_location, edge_spacing_x,
edge_spacing_y) = cropPadImage(bbox_curr_shift, curr_img)
bbox_curr_gt = BoundingBox(currbb[0], currbb[1], currbb[2], currbb[3])
bbox_gt_recentered = BoundingBox(0, 0, 0, 0)
bbox_gt_recentered = bbox_curr_gt.recenter(rand_search_location,
edge_spacing_x,
edge_spacing_y,
bbox_gt_recentered)
# get larger context
bbox_curr_shift.kContextFactor = 4
(rand_search_region_x2, rand_search_location_x2, edge_spacing_x_x2,
edge_spacing_y_x2) = cropPadImage(bbox_curr_shift, curr_img)
curr_sample['image'] = rand_search_region
curr_sample['image_x2'] = rand_search_region_x2
curr_sample['bb'] = bbox_gt_recentered.get_bb_list()
# additional options for visualization
opts_curr['edge_spacing_x'] = edge_spacing_x
opts_curr['edge_spacing_y'] = edge_spacing_y
opts_curr['search_location'] = rand_search_location
opts_curr['search_region'] = rand_search_region
# build prev sample
prev_sample = self.get_orig_sample(idx, 0)
prev_sample_x2 = self.get_orig_sample(idx, 0)
prev_sample, opts_prev = crop_sample(prev_sample)
prev_sample_x2, opts_prev_x2 = crop_sample(prev_sample_x2, 4)
prev_sample['image_x2'] = prev_sample_x2['image']
# scale
scale = Rescale((self.input_size, self.input_size))
scaled_curr_obj = scale(curr_sample, opts_curr)
scaled_prev_obj = scale(prev_sample, opts_prev)
training_sample = {
'previmg': scaled_prev_obj['image'],
'currimg': scaled_curr_obj['image'],
'previmg_x2': scaled_prev_obj['image_x2'],
'currimg_x2': scaled_curr_obj['image_x2'],
'currbb': scaled_curr_obj['bb']
}
return training_sample, opts_curr
def make_transformed_samples(dataset, args):
'''
Given a dataset, it picks a random sample from it and returns a batch
of (kGeneratedExamplesPerImage+1) samples. The batch contains true sample
from dataset and kGeneratedExamplesPerImage samples, which are created
artifically with augmentation by GOTURN smooth motion model.
'''
idx = np.random.randint(dataset.len, size=1)[0]
# unscaled original sample (single image and bb)
orig_sample = dataset.get_orig_sample(idx)
# cropped scaled sample (two frames and bb)
true_sample, _ = dataset.get_sample(idx)
true_tensor = transform(true_sample)
x1_batch = torch.Tensor(kGeneratedExamplesPerImage + 1, 3, input_size,
input_size)
x2_batch = torch.Tensor(kGeneratedExamplesPerImage + 1, 3, input_size,
input_size)
x1_batch_x2 = torch.Tensor(kGeneratedExamplesPerImage + 1, 3,
input_size * 2, input_size * 2)
x2_batch_x2 = torch.Tensor(kGeneratedExamplesPerImage + 1, 3,
input_size * 2, input_size * 2)
y_batch = torch.Tensor(kGeneratedExamplesPerImage + 1, 4)
# initialize batch with the true sample
x1_batch[0] = true_tensor['previmg']
x2_batch[0] = true_tensor['currimg']
x1_batch_x2[0] = true_tensor['previmg_x2']
x2_batch_x2[0] = true_tensor['currimg_x2']
y_batch[0] = true_tensor['currbb']
scale = Rescale((input_size, input_size))
for i in range(kGeneratedExamplesPerImage):
sample = orig_sample
# unscaled current image crop with box
curr_sample, opts_curr = shift_crop_training_sample(sample, bb_params)
# unscaled previous image crop with box
prev_sample, opts_prev = crop_sample(sample)
prev_sample_x2, opts_prev_x2 = crop_sample(sample, contextFactor=4)
prev_sample['image_x2'] = prev_sample_x2['image']
scaled_curr_obj = scale(curr_sample, opts_curr)
scaled_prev_obj = scale(prev_sample, opts_prev)
training_sample = {
'previmg': scaled_prev_obj['image'],
'currimg': scaled_curr_obj['image'],
'previmg_x2': scaled_prev_obj['image_x2'],
'currimg_x2': scaled_curr_obj['image_x2'],
'currbb': scaled_curr_obj['bb']
}
sample = transform(training_sample)
x1_batch[i + 1] = sample['previmg']
x2_batch[i + 1] = sample['currimg']
x1_batch_x2[i + 1] = sample['previmg_x2']
x2_batch_x2[i + 1] = sample['currimg_x2']
y_batch[i + 1] = sample['currbb']
return x1_batch, x2_batch, x1_batch_x2, x2_batch_x2, y_batch
def _get_sample(self, idx):
"""
Returns cropped previous and current frame at the previous predicted
location. Note that the images are scaled to (224,224,3).
"""
prev = self.img[idx][0]
curr = self.img[idx][1]
prevbb = self.prev_rect
prev_sample, opts_prev = crop_sample({'image': prev, 'bb': prevbb})
curr_sample, opts_curr = crop_sample({'image': curr, 'bb': prevbb})
curr_img = self.scale(curr_sample, opts_curr)['image']
prev_img = self.scale(prev_sample, opts_prev)['image']
sample = {'previmg': prev_img, 'currimg': curr_img}
self.curr_img = curr
self.opts = opts_curr
return sample
def _get_sample(self):
"""
Returns cropped previous and current frame at the previous predicted
location. Note that the images are scaled to (224,224,3).
"""
prev = self.prev_img
curr = self.curr_img
prevbb = self._last_bbox
prev_sample, opts_prev = crop_sample({'image': prev, 'bb': prevbb})
curr_sample, opts_curr = crop_sample({'image': curr, 'bb': prevbb})
prev_img = bgr2rgb(self.scale(prev_sample, opts_prev)['image'])
curr_img = bgr2rgb(self.scale(curr_sample, opts_curr)['image'])
sample = {'previmg': prev_img, 'currimg': curr_img}
self.curr_img = curr
self.opts = opts_curr
return sample
def make_transformed_samples(dataset, args):
idx = np.random.randint(dataset.len, size=1)[0]
# unscaled original sample (single image and bb)
orig_sample = dataset.get_orig_sample(idx)
# cropped scaled sample (two frames and bb)
true_sample, _ = dataset.get_sample(idx)
true_tensor = transform(true_sample)
x1_batch = torch.Tensor(kGeneratedExamplesPerImage + 1, 3, input_size,
input_size)
x2_batch = torch.Tensor(kGeneratedExamplesPerImage + 1, 3, input_size,
input_size)
y_batch = torch.Tensor(kGeneratedExamplesPerImage + 1, 4)
# initialize batch with the true sample
x1_batch[0, :, :, :] = true_tensor['previmg']
x2_batch[0, :, :, :] = true_tensor['currimg']
y_batch[0, :] = true_tensor['currbb']
scale = Rescale((input_size, input_size))
for i in range(kGeneratedExamplesPerImage):
sample = orig_sample
# unscaled current image crop with box
curr_sample, opts_curr = shift_crop_training_sample(sample, bb_params)
# unscaled previous image crop with box
prev_sample, opts_prev = crop_sample(sample)
scaled_curr_obj = scale(curr_sample, opts_curr)
scaled_prev_obj = scale(prev_sample, opts_prev)
training_sample = {
'previmg': scaled_prev_obj['image'],
'currimg': scaled_curr_obj['image'],
'currbb': scaled_curr_obj['bb']
}
sample = transform(training_sample)
x1_batch[i + 1, :, :, :] = sample['previmg']
x2_batch[i + 1, :, :, :] = sample['currimg']
y_batch[i + 1, :] = sample['currbb']
return x1_batch, x2_batch, y_batch
def get_sample(self, idx):
"""
Returns sample without transformation for visualization.
Sample consists of resized previous and current frame with target
which is passed to the network. Bounding box values are normalized
between 0 and 1 with respect to the target frame and then scaled by
factor of 10.
"""
sample = self.get_orig_sample(idx)
# unscaled current image crop with box
curr_sample, opts_curr = shift_crop_training_sample(
sample, self.bb_params)
# unscaled previous image crop with box
prev_sample, opts_prev = crop_sample(sample)
scale = Rescale((self.sz, self.sz))
scaled_curr_obj = scale(curr_sample, opts_curr)
scaled_prev_obj = scale(prev_sample, opts_prev)
training_sample = {
'previmg': scaled_prev_obj['image'],
'currimg': scaled_curr_obj['image'],
'currbb': scaled_curr_obj['bb']
}
return training_sample, opts_curr
