def run_on_block(c_im, h_im, padding=0):
"""Run complete model on the block c_im and its corresponding height block h_im."""
if c_im.mean() <= 2: # black part
return [], np.array([])
green_centers = proc.green_hotspots(
c_im, sigma=sigma, padding=padding) # run green region proposer
dark_centers = proc.dark_hotspots(
c_im, sigma=sigma, padding=padding) # run dark region proposer
c_coords = np.concatenate((green_centers, dark_centers))
c_rects, h_rects = create_boxes(c_coords)
c_crops, h_crops = fill_data_tensor(c_im, h_im, c_rects, h_rects)
predictions = box_model.predict([c_crops, h_crops],
verbose=1) # run classification model
sorted_predictions = proc.multi_class_sort(c_rects, predictions)
output = [([], [], []) for k in range(len(sorted_predictions))]
for (k, (rects, probs)) in enumerate(sorted_predictions):
if len(rects) > 0:
rects, probs = proc.non_max_suppression(rects,
probs=probs,
t=overlap_threshold)
masks = proc.get_masks(rects, c_im, mask_models[k], verbose=1)
output[k] = (rects, probs, masks)
new_output = [([], []) for k in range(len(output))]
for (k, (rects, probs, masks)) in enumerate(output):
try:
if filter_empty_masks:
rects, probs, masks = proc.discard_empty(rects,
probs,
masks,
t=crop_size_threshold)
if filter_disjoint:
masks = proc.remove_unconnected_components(masks)
if recenter:
rects, altered = proc.recenter_boxes(
rects, masks, d=center_distance) # indeces of moved boxes
new_masks = proc.get_masks(
rects[altered], c_im, mask_models[k],
verbose=1) # compute new masks of moved boxes
if filter_disjoint:
new_masks = proc.remove_unconnected_components(new_masks)
masks[altered] = new_masks # set new masks
if filter_empty_masks:
rects, probs, masks = proc.discard_empty(
rects, probs, masks, t=crop_size_threshold)
cnts = proc.find_contours(rects, masks)
ctrs = proc.find_centroids(rects, masks)
new_output[k] = (cnts, ctrs)
except:
print('No {} lettuce found'.format(['drk', 'grn'][k]))
return new_output
def run_on_block(c_im, h_im, padding=0, get_background=False):
"""Run complete model on the block c_im and its corresponding height block h_im."""
if c_im.mean() <= 1e-6: # black part
raise IndexError
c_coords = proc.green_hotspots(c_im, sigma=sigma,
padding=padding) # run region proposer
c_rects, h_rects = create_boxes(c_coords, box_size)
input_RGB, input_DEM = fill_data_tensor(c_im, h_im, c_rects, h_rects)
# --------
input_RGB = proc.apply_preprocessing(input_RGB, function=proc.cielab)
# --------
predictions, masks = network.predict([input_RGB, input_DEM],
verbose=1) # run classification model
masks = masks[..., 0]
crop_idxs = proc.get_class_idxs(predictions, 1)
boxes, [confidence, masks
] = c_rects[crop_idxs], [predictions[crop_idxs], masks[crop_idxs]]
boxes, [confidence,
masks] = proc.non_max_suppression(boxes,
other=[confidence, masks],
t=overlap_threshold)
masks = proc.get_hard_masks(masks)
if filter_empty_masks:
masks, boxes, [confidence] = proc.discard_empty(masks,
boxes,
other=[confidence],
t=crop_size_threshold)
contours = proc.find_contours(boxes, masks)
centroids = proc.find_centroids(boxes, masks)
if get_background:
background_boxes, background_confidence = proc.get_class(
c_rects, predictions, 0)
return [contours, centroids, boxes,
confidence], [background_boxes, background_confidence]
else:
return [contours, centroids, boxes, confidence], [[], []]
def run_on_block(c_im, h_im, padding=0, get_background=False): """Run complete model on the block c_im and its corresponding height block h_im.""" if c_im.mean() <= 2: # black part raise IndexError c_coords = proc.green_hotspots(c_im, sigma=sigma, padding=padding) # run region proposer c_rects, h_rects = create_boxes(c_coords) c_crops, h_crops = fill_data_tensor(c_im, h_im, c_rects, h_rects) predictions = box_model.predict([c_crops, h_crops], verbose=1) # run classification model boxes, confidence = proc.get_class(c_rects, predictions, 1) boxes, [confidence] = proc.non_max_suppression(boxes, other=[confidence], t=overlap_threshold) masks = proc.get_masks(boxes, c_im, mask_model, verbose=1) # compute masks for each box if filter_empty_masks: masks, [boxes, confidence] = proc.discard_empty(masks, other=[boxes, confidence], t=crop_size_threshold) if filter_disjoint: masks = proc.remove_unconnected_components(masks) if recenter: boxes, altered = proc.recenter_boxes(boxes, masks, d=center_distance) # indeces of moved boxes new_masks = proc.get_masks(boxes[altered], c_im, mask_model, verbose=1) # compute new masks of moved boxes if filter_disjoint: new_masks = proc.remove_unconnected_components(new_masks) masks[altered] = new_masks # set new masks if filter_empty_masks: boxes, confidence, masks = proc.discard_empty(boxes, confidence, masks, t=crop_size_threshold) contours = proc.find_contours(boxes, masks) centroids = proc.find_centroids(boxes, masks) if get_background: background_boxes, background_confidence = proc.get_class(c_rects, predictions, 0) return contours, centroids, confidence, boxes, background_boxes, background_confidence else: return contours, centroids, confidence, boxes
def run_on_block(self, rgb_block, dem_block, get_background=False):
"""Run detection algorithm on RGB block and its corresponding DEM block.
A short summary of the detection algorithm:
* First generate RoI's, and put boxes of a fixed size at these locations.
* Feed the data in the boxes through a combined classification and
masking network.
* Sort the results into crops and background.
* Apply post-processing to crop results, like non-max-suppression to
discard overlapping boxes, and clean up masks.
* Convert masks to contours and centroids.
Arguments
---------
rgb_block : (?,?,3) block
RGB data block.
dem_block : (?,?) block
DEM data block corresponding to RGB block.
get_background : bool, optional
If set to True, all boxes that contain background are stored
and returned
Returns
-------
crop_output : list of length 4
List containing contours, centroids, boxes and confidence scores
bg_output : list of length 2
List containing background boxes and confidence scores. If
get_background==False, this is a list containing two empty lists.
Raises
------
IndexError : if the input block is completely black (no data).
"""
if rgb_block.mean() <= 1e-6:
raise IndexError('This block contains no data.')
rgb_coords = processing.green_hotspots(rgb_block, sigma=self.Settings.sigma, padding=self.Settings.box_size) # run region proposer
rgb_boxes, dem_boxes = self.create_boxes(rgb_coords, self.Settings.box_size)
rgb_input_tensor, dem_input_tensor = self.fill_data_tensors(rgb_block, dem_block, rgb_boxes, dem_boxes)
predictions, masks = self.network.predict([rgb_input_tensor, dem_input_tensor], verbose=1) # run classification model
masks = masks[...,0]
output = dict()
for class_idx in range(1, self.num_classes):
cls_idxs = processing.get_class_idxs(predictions, class_idx)
cls_boxes, [cls_confidence, cls_masks] = rgb_boxes[cls_idxs], [predictions[cls_idxs], masks[cls_idxs]]
cls_boxes, [cls_confidence, cls_masks] = processing.non_max_suppression(cls_boxes, other=[cls_confidence, cls_masks], t=self.Settings.overlap_threshold)
cls_masks = processing.get_hard_masks(cls_masks)
cls_masks, cls_boxes, [cls_confidence] = processing.discard_empty(cls_masks, cls_boxes, other=[cls_confidence], t=self.Settings.crop_size_threshold)
cls_contours = processing.find_contours(cls_boxes, cls_masks)
cls_centroids = processing.find_centroids(cls_boxes, cls_masks)
output[class_idx] = {'contours' : cls_contours,
'centroids' : cls_centroids,
'boxes' : cls_boxes,
'confidence' : cls_confidence}
if get_background:
bg_idxs = processing.get_class_idxs(predictions, 0)
background_boxes, background_confidence = rgb_boxes[bg_idxs], predictions[bg_idxs]
return output, [background_boxes, background_confidence]
else:
return output, [[],[]]