def rssd_predictions_to_ship_mask(self, boxes, labels):
boxes, scores = self.box_coder.decode(boxes, labels)
mask = np.zeros(
self.image_size, dtype=np.uint16
) # Just in case we have more than 255 ships in one image
for i, bbox in enumerate(boxes):
visualize_rbbox(mask,
bbox, (i + 1, i + 1, i + 1),
thickness=cv2.FILLED)
return mask
def test_draw_sersdd_bboxes():
box_coder = RRNBoxCoder(768, 768)
anchors = box_coder._get_anchor_wht()
n = len(box_coder.anchor_boxes)
print('Total number of anchor boxes in SERSSDBoxCoder',
len(box_coder.anchor_boxes))
for i, (wht_fm,
fm_size) in enumerate(zip(anchors, box_coder.feature_map_sizes)):
image = np.zeros((box_coder.image_height, box_coder.image_width, 3),
dtype=np.uint8)
grid_size = box_coder.input_size / fm_size
fm_w, fm_h = int(fm_size[0]), int(fm_size[1])
xy = meshgrid_numpy(fm_w, fm_h).astype(np.float32) + 0.5
xy = (xy * grid_size).reshape(-1, 2)
for p in xy:
p = p.astype(int)
cv2.circle(image, tuple(p), 2, (200, 200, 200), cv2.FILLED)
for wht in wht_fm:
a = math.sqrt(wht[0] * wht[1])
w = int(a)
h = int(a)
cv2.rectangle(image, (p[0] - w // 2, p[1] - h // 2),
(p[0] + w // 2, p[1] + h // 2), (100, 100, 100))
# rbox = [p[0], p[1], wht[0], wht[1], wht[2]]
# visualize_rbbox(image, rbox, (i * 28, 255, 0), thickness=1)
for wht in wht_fm:
rbox = [
box_coder.image_height // 2, box_coder.image_width // 2,
wht[0], wht[1], wht[2]
]
visualize_rbbox(image, rbox, (i * 28, 255, 0), thickness=1)
cv2.imshow("Image" + str(i), image)
cv2.waitKey(-1)
def predict_as_csv(self, dataset, batch_size=1, workers=0):
self.eval()
dataloader = DataLoader(dataset,
batch_size=batch_size,
num_workers=workers)
image_ids = []
rles = []
for image, y in tqdm(dataloader, total=len(dataloader)):
image = image.cuda(non_blocking=True)
pred_boxes, pred_labels = self(image)
for image_id, boxes, scores in zip(y[ID_KEY], to_numpy(pred_boxes),
to_numpy(pred_labels)):
boxes, _, scores = dataset.box_coder.decode(boxes, scores)
if len(boxes):
mask = np.zeros((768, 768), dtype=np.uint8)
# First, we resterize all masks
for i, rbox in enumerate(boxes):
visualize_rbbox(mask,
rbox,
color=(i + 1, i + 1, i + 1),
thickness=cv2.FILLED)
# Second, we do rle encoding. This prevents assigning same pixel to multiple instances
for i, rbox in enumerate(boxes):
rle = rle_encode(mask == (i + 1))
image_ids.append(image_id)
rles.append(rle)
else:
image_ids.append(image_id)
rles.append(None)
return pd.DataFrame.from_dict({
'ImageId': image_ids,
'EncodedPixels': rles
})
def test_rssd_encoding():
id = 'dc6adaf6b.jpg'
groundtruth = pd.read_csv(
os.path.join(data_dir, 'train_ship_segmentations_v2.csv'))
image = read_train_image(id, data_dir)
rle = get_rle_from_groundtruth(groundtruth, id)
label_image = rle2instancemask(rle)
# Test what happens if we rotate
# image = np.rot90(image).copy()
# label_image = np.rot90(label_image).copy()
rbboxes = instance_mask_to_rbboxes(label_image)
print('Instances', rbboxes)
labels = np.zeros(len(rbboxes), dtype=np.intp)
box_coder = RSSDBoxCoder(768, 778)
loc_targets, cls_targets = box_coder.encode(rbboxes, labels)
print(loc_targets.shape, cls_targets.shape)
cls_targets_one_hot = np.eye(2)[cls_targets]
print(cls_targets_one_hot.shape)
dec_boxes, dec_labels, dec_scores = box_coder.decode(
loc_targets, cls_targets_one_hot)
print(dec_boxes)
for bbox in dec_boxes:
visualize_rbbox(image, bbox, (255, 0, 0), thickness=3)
for bbox in rbboxes:
visualize_rbbox(image, bbox, (0, 255, 0), thickness=1)
cv2.imshow('a', image)
cv2.waitKey(-1)
def test_draw_rsdd_bboxes():
box_coder = RSSDBoxCoder(768, 768)
anchors = box_coder._get_anchor_wht()
n = len(FPNSSDBoxCoder().anchor_boxes)
print('Total number of anchor boxes SSD', len(box_coder.anchor_boxes))
n = len(box_coder.anchor_boxes)
print('Total number of anchor boxes in RSSD', len(box_coder.anchor_boxes))
for i, wht_fm in enumerate(anchors):
image = np.zeros((box_coder.image_height, box_coder.image_width, 3),
dtype=np.uint8)
for wht in wht_fm:
rbox = [
box_coder.image_height // 2, box_coder.image_width // 2,
wht[0], wht[1], wht[2]
]
visualize_rbbox(image, rbox, (i * 28, 255, 0), thickness=1)
cv2.imshow("Image" + str(i), image)
cv2.waitKey(-1)
def test_mask():
# id = 'aed55df18.jpg'
id = 'dc6adaf6b.jpg'
groundtruth = pd.read_csv(
os.path.join(data_dir, 'train_ship_segmentations_v2.csv'))
image = read_train_image(id, data_dir)
rle = get_rle_from_groundtruth(groundtruth, id)
label_image = rle2instancemask(rle)
mask = instance_mask_to_fg_and_edge(label_image)
dummy = np.zeros((label_image.shape[0], label_image.shape[1], 1),
dtype=np.uint8)
mask = np.concatenate([mask * 255, dummy], axis=2)
cv2.imshow('mask (no edge)', (label_image > 0).astype(np.uint8) * 255)
cv2.imshow('mask', mask)
# cv2.waitKey(-1)
rbboxes = instance_mask_to_rbboxes(label_image)
for bbox in rbboxes:
visualize_rbbox(image, bbox, (0, 255, 0))
cv2.imshow('a', image)
cv2.waitKey(-1)
def test_rrssd_box_coder_synthetic():
label_image = np.zeros((768, 768), dtype=np.uint8)
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((100, 100), (100, 20), 0)),
1).astype(int), (1, 1, 1))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((200, 100), (100, 20), 45)),
1).astype(int), (2, 2, 2))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((100, 200), (100, 20), 90)),
1).astype(int), (3, 3, 3))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((200, 200), (100, 20), 135)),
1).astype(int), (4, 4, 4))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((100 + 200, 100), (20, 100), 0)),
1).astype(int), (5, 5, 5))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((200 + 200, 100), (20, 100), 45)),
1).astype(int), (6, 6, 6))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((100 + 200, 200), (20, 100), 90)),
1).astype(int), (7, 7, 7))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((200 + 200, 200), (20, 100), 135)),
1).astype(int), (8, 8, 8))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((100, 105 + 200), (100, 20), 45. / 2)),
1).astype(int), (9, 9, 9))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((200, 100 + 200), (16, 4), 49)),
1).astype(int), (10, 10, 10))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((200, 100 + 210), (100, 20), 49)),
1).astype(int), (11, 11, 11))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((200, 200 + 200), (100, 20), 165)),
1).astype(int), (12, 12, 12))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((300, 300), (10, 6), 49)),
1).astype(int), (13, 13, 13))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((300 + 50, 300), (200, 40), 90)),
1).astype(int), (14, 14, 14))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((300 + 70, 300), (100, 20), 90)),
1).astype(int), (15, 15, 15))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((500, 500), (2, 3), 9)), 1).astype(int),
(16, 16, 16))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((510, 500), (2, 3), 19)), 1).astype(int),
(17, 17, 17))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((540, 500), (4, 6), 29)), 1).astype(int),
(18, 18, 18))
cv2.fillConvexPoly(
label_image,
np.expand_dims(cv2.boxPoints(((560, 500), (2, 3), 39)), 1).astype(int),
(19, 19, 19))
image = (label2rgb(label_image, bg_label=0) * 255).astype(np.uint8)
# Test what happens if we rotate
# image = np.rot90(image).copy()
# label_image = np.rot90(label_image).copy()
rbboxes = instance_mask_to_rbboxes(label_image)
print(rbboxes)
labels = np.zeros(len(rbboxes), dtype=np.intp)
box_coder = RRNBoxCoder(768, 768)
loc_targets, cls_targets, anchors = box_coder.encode([], [],
return_anchors=True)
loc_targets, cls_targets, anchors = box_coder.encode(rbboxes,
labels,
return_anchors=True)
print(loc_targets.shape, cls_targets.shape, anchors.shape)
print('Object anchors', (cls_targets > 0).sum())
print('Background anchors', (cls_targets == 0).sum())
print('Ignore anchors', (cls_targets == -1).sum())
# cls_targets = np.expand_dims(cls_targets)
# print(cls_targets_one_hot.shape)
dec_boxes, dec_scores = box_coder.decode(loc_targets, cls_targets)
print(dec_boxes)
print('Total anchors', len(anchors))
for bbox in dec_boxes:
visualize_rbbox(image, bbox, (255, 0, 255), thickness=3)
for bbox in rbboxes:
visualize_rbbox(image, bbox, (0, 255, 0), thickness=1)
for bbox in anchors:
visualize_rbbox(image, bbox, (255, 255, 255), thickness=1)
cv2.imshow('overlays', image)
cv2.waitKey(-1)