def coco_write_anns_to_file(coco, img_id):
"""
Given an image id $img_id, write out the annotations in the
mAP format to $MAP_GROUND_TRUTH_DIR/$image_id.txt
"""
anns = get_image_annotations(coco, img_id)
fname = '%s/%012d.%s' % (MAP_GROUND_TRUTH_DIR, img_id, 'txt')
write_anns_to_file(anns, fname)
def test_single_abs_ann_to_rect_mask():
import matplotlib.pyplot as plt
from pycocotools.coco import COCO
from file_management import ANNOTATION_FILE, get_downloaded_ids
from annotations import get_image_annotations, plot_annotations
coco = COCO(ANNOTATION_FILE)
img_ids = get_downloaded_ids()
anns = [get_image_annotations(coco, img) for img in img_ids]
#view_img = lambda index: plot_annotations(img_ids[index], anns[index])
ann = anns[0][0]
plt.figure(2)
fig, ax = plt.subplots()
ax.imshow(single_abs_ann_to_rect_mask(640, 420, ann['bbox']))
plt.figure(1)
plot_annotations(img_ids[0], [ann])
def test_abs_anns_to_heatmap():
import matplotlib.pyplot as plt
from matplotlib import cm
from pycocotools.coco import COCO
from file_management import ANNOTATION_FILE, get_downloaded_ids
from annotations import get_image_annotations, plot_annotations
coco = COCO(ANNOTATION_FILE)
img_ids = get_downloaded_ids()
anns = [get_image_annotations(coco, img) for img in img_ids]
plt.figure(0)
ann = anns[0]
heatmap = abs_anns_to_heatmap(640, 640, ann)
plt.figure(2)
fig, ax = plt.subplots()
ax.imshow(heatmap) #, cmap=cm.jet)
plt.figure(1)
plot_annotations(img_ids[0], ann)
def get_y_true(coco, bounding_box_count, cell_width_px, cell_height_px,
img_id):
# Force bounding_box_count to 1 - See NOTE in above documentation
bounding_box_count = 1
# These should probably get moved to the parameters
DEFAULT_VALUE = 0
NO_OBJECT_WEIGHT = 0
HAS_OBJECT_WEIGHT = 1
# Position of the various training parameters along the last dimension
# of the output data from the neural network
POS_OBJ_SCORE = 0
POS_BOX_CENTER_X = 1
POS_BOX_CENTER_Y = 2
POS_BOX_WIDTH = 3
POS_BOX_HEIGHT = 4
annotations = get_image_annotations(coco, img_id)
# img = coco.loadImgs([img_id])[0]
# cell_x_count, how many cells are on horizontal direction, cell_y_count,
# how many cells are on vertical direction
cell_x_count = ceil(PADDED_SIZE / cell_width_px)
cell_y_count = ceil(PADDED_SIZE / cell_height_px)
# 5 parameters to each bounding box: Probability, X pos, Y pos, Width, Height
y_true = np.full((cell_y_count, cell_x_count, bounding_box_count * 5),
DEFAULT_VALUE)
y_true = y_true.astype('float32')
# Set all object probabilities to NO_OBJECT_WEIGHT
if DEFAULT_VALUE != NO_OBJECT_WEIGHT:
y_true[..., ..., POS_OBJ_SCORE::5] = NO_OBJECT_WEIGHT
for annotation in annotations:
# Calculate the cell that the annotation should match
bounding_box = annotation['bbox']
# print("[DEBUG]", bounding_box)
abs_ul_x = bounding_box[0]
abs_ul_y = bounding_box[1]
width = bounding_box[2]
height = bounding_box[3]
# Find the center of the box in terms of the whole image
# These values are purposely floats to keep as much information as
# possible about the center of the image
abs_center_x = abs_ul_x + width / 2
abs_center_y = abs_ul_y + height / 2
# Calculate the cell the bounding box is centered in
cell_x_pos = floor(abs_center_x / cell_width_px)
cell_y_pos = floor(abs_center_y / cell_height_px)
# Find the center of the box relative to the corner of the cell:
# ...And put it in terms of the cell size
rel_center_x = (abs_center_x -
(cell_x_pos * cell_width_px)) / cell_width_px
rel_center_y = (abs_center_y -
(cell_y_pos * cell_height_px)) / cell_height_px
# Find the size of the bounding box relative to the cell
rel_width = (width / cell_width_px) / 10 #hard code
rel_height = (height / cell_height_px) / 10 #hard code
# TODO: Move to handling more than one bounding box
# if y_true[cell_y_pos, cell_x_pos, POS_OBJ_SCORE] != NO_OBJECT_WEIGHT:
# logging.warn("Image %d has multiple bounding boxes in cell (%d,%d)" % (
# img_id,
# cell_x_pos,
# cell_y_pos
# ))
# Set values for the training data
y_true[cell_y_pos, cell_x_pos, POS_BOX_CENTER_X] = rel_center_x
y_true[cell_y_pos, cell_x_pos, POS_BOX_CENTER_Y] = rel_center_y
y_true[cell_y_pos, cell_x_pos, POS_BOX_WIDTH] = rel_width
y_true[cell_y_pos, cell_x_pos, POS_BOX_HEIGHT] = rel_height
y_true[cell_y_pos, cell_x_pos, POS_OBJ_SCORE] = HAS_OBJECT_WEIGHT
# print("[DEBUG]", y_true[cell_y_pos, cell_x_pos, :])
return y_true
#!/usr/bin/env python3 -i from pycocotools.coco import COCO from file_management import ANNOTATION_FILE, get_downloaded_ids from annotations import get_image_annotations, plot_annotations coco = COCO(ANNOTATION_FILE) img_ids = get_downloaded_ids() anns = [get_image_annotations(coco, img) for img in img_ids] view_img = lambda index: plot_annotations(img_ids[index], anns[index])