def pull_item(self, idx):
file = self.files[self.index[idx]]
file = file.replace('\\', '/')
image = file.replace('.json', '.jpg')
json_data = json.loads(open(file).read())
image = io.imread(image)
image, window, scale, padding, crop = resize_image_fixed_size(
image, self.image_size)
label_ix = []
box = []
for s in json_data['shapes']:
label_ix.append(self.class_to_idx[s['label']])
box.append([
s['points'][0][0], s['points'][0][1], s['points'][2][0],
s['points'][2][1]
])
box = np.asarray(box)
box = box * scale
box[:, 0] = box[:, 0] + padding[1][0]
box[:, 1] = box[:, 1] + padding[0][0]
box[:, 2] = box[:, 2] + padding[1][1]
box[:, 3] = box[:, 3] + padding[0][1]
box = box / np.asarray([
self.image_size[0], self.image_size[1], self.image_size[0],
self.image_size[1]
])
return image, box, label_ix
def run():
batch_size = 1
image = tf.placeholder(dtype=tf.float32, shape=(batch_size, 512, 512, 3))
global_step = tf.train.get_or_create_global_step()
out_put = resnet50.fpn(image)
out_put = tf.nn.sigmoid(out_put)
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver.restore(sess, 'drive/model.ckpt-53581')
for x in glob.glob(
'/media/dsl/20d6b919-92e1-4489-b2be-a092290668e4/BDD100K/bdd100k/images/100k/val/*.*'
):
ig = io.imread(x)
org, window, scale, padding, crop = utils.resize_image_fixed_size(
ig, config.image_size)
ig = org - [123.15, 115.90, 103.06]
ig = np.expand_dims(ig, 0)
fd = {image: ig}
o_put = sess.run(out_put, feed_dict=fd)
plt.subplot(221)
plt.title('step1')
plt.imshow(o_put[0, :, :, 0], aspect="auto", cmap='gray')
plt.subplot(222)
plt.title('step2')
plt.imshow(o_put[0, :, :, 1], aspect="auto", cmap='gray')
plt.subplot(223)
plt.title('org')
plt.imshow(org, aspect="auto")
plt.show()
def pull_item(self, idx):
item_data = self.data[idx]
image_name = item_data['name']
labels = item_data['labels']
label_ix = []
box = []
ig_data = io.imread(os.path.join(self.image_dr, image_name))
for ll in labels:
category_name = ll['category']
if self.class_mapix.get(category_name, None) is not None:
box2d = ll['box2d']
label_ix.append(self.class_mapix[category_name])
box.append(
[box2d['x1'], box2d['y1'], box2d['x2'], box2d['y2']])
box = np.asarray(box)
if self.is_crop:
ig, box, label_ix = utils.crop_image_with_box(
ig_data, self.image_size, box, label_ix)
else:
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
ig_data, self.image_size)
box = box * scale
box[:, 0] = box[:, 0] + padding[1][0]
box[:, 1] = box[:, 1] + padding[0][0]
box[:, 2] = box[:, 2] + padding[1][1]
box[:, 3] = box[:, 3] + padding[0][1]
box = box / np.asarray([
self.image_size[1], self.image_size[0], self.image_size[1],
self.image_size[0]
])
return ig, box, label_ix
def pull_item(self, idx):
json_pth = self.data[idx]
image_pth = json_pth.replace('.json', '.png')
ig_data = io.imread(image_pth)[:, :, 0:3]
instance_masks = np.zeros(
shape=[self.image_size[0], self.image_size[1], 3])
js_data = json.loads(open(json_pth).read())
msk = []
ids = []
for b in js_data['boundary']:
label = b['correctionType']
if label == 'land':
points = b['points']
p = []
for pp in points:
p.append([pp['pix_x'], pp['pix_y']])
poly_land = np.zeros(
shape=[self.image_size[0], self.image_size[1], 3],
dtype=np.uint8)
cv2.fillPoly(poly_land, np.asarray([p], np.int),
(255, 255, 255))
msk.append(poly_land[:, :, 0:1])
ids.append(0)
if len(msk) > 1:
msk = np.concatenate(msk, axis=2)
elif len(msk) == 1:
msk = msk[0]
else:
return None
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
ig_data, self.image_size)
msk = utils.resize_mask(msk, scale, padding, crop)
if random.randint(0, 1) == 2:
ag = self.aug.to_deterministic()
ig_data = ag.augment_image(ig_data)
msk = ag.augment_image(msk)
box = utils.extract_bboxes(msk)
ids = np.asarray(ids)
mask = np.sum(msk, 2)
mask[np.where(mask > 255)] = 255
box = box / np.asarray([
self.image_size[0], self.image_size[1], self.image_size[0],
self.image_size[1]
])
return ig, box, ids, mask
def pull_item(self, idx):
item_data = self.data[idx]
image_name = item_data['name']
labels = item_data['labels']
label_ix = []
box = []
ig_data = io.imread(os.path.join(self.image_dr, image_name))
direct = np.zeros(shape=ig_data.shape, dtype=np.uint8)
alter = np.zeros(shape=ig_data.shape, dtype=np.uint8)
for ll in labels:
category_name = ll['category']
if category_name == 'drivable area':
if ll['attributes']['areaType'] == 'direct':
pts = ll['poly2d']
for x in pts:
cv2.fillPoly(direct,
[np.asarray(x['vertices'], np.int)],
(255, 255, 255))
elif ll['attributes']['areaType'] == 'alternative':
pts = ll['poly2d']
for x in pts:
cv2.fillPoly(alter,
[np.asarray(x['vertices'], np.int)],
(255, 255, 255))
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
ig_data, self.image_size)
direct, window, scale, padding, crop = utils.resize_image_fixed_size(
direct, self.image_size)
alter, window, scale, padding, crop = utils.resize_image_fixed_size(
alter, self.image_size)
labels = np.zeros(shape=(ig.shape[0], ig.shape[1], 2), dtype=np.uint8)
labels[:, :, 0] = direct[:, :, 0]
labels[:, :, 1] = alter[:, :, 0]
return ig, labels
def pull_item(self, idx):
item_data = self.data[idx]
image_name = item_data['name']
labels = item_data['labels']
label_ix = []
box = []
ig_data = io.imread(os.path.join(self.image_dr, image_name))
instance_masks = []
cls_ids = []
for ll in labels:
category_name = ll['category']
if category_name == 'drivable area':
if ll['attributes']['areaType'] == 'direct':
pts = ll['poly2d']
for x in pts:
direct = np.zeros(shape=ig_data.shape, dtype=np.uint8)
cv2.fillPoly(direct,
[np.asarray(x['vertices'], np.int)],
(255, 255, 255))
cls_ids.append(0)
instance_masks.append(direct[:, :, 0])
elif ll['attributes']['areaType'] == 'alternative':
pts = ll['poly2d']
for x in pts:
alter = np.zeros(shape=ig_data.shape, dtype=np.uint8)
cv2.fillPoly(alter,
[np.asarray(x['vertices'], np.int)],
(255, 255, 255))
cls_ids.append(1)
instance_masks.append(alter[:, :, 0])
mask = np.asarray(instance_masks)
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
ig_data, self.image_size)
if len(labels) == 0:
return
mask = np.transpose(mask, axes=[1, 2, 0])
mask = utils.resize_mask(mask, scale, padding, crop)
image, mask = coco_handler.aug(ig, mask)
boxes = utils.extract_bboxes(mask)
mask = utils.minimize_mask(boxes,
mask,
mini_shape=(self.mask_shape,
self.mask_shape))
#boxes = boxes /np.asarray([self.image_size[1], self.image_size[0],self.image_size[1], self.image_size[0]])
return ig, cls_ids, boxes, mask
def pull_item(self, index):
img_id = os.path.join(self.image_path, self.images[index])
boxes = self.boxes[index]
labels = np.ones(len(boxes)) * 0
if len(boxes) == 0:
return None, None, None
img = io.imread(img_id)
box = np.asarray(boxes)
if True:
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
img, self.image_size)
if len(labels) == 0:
return ig, box, labels
box = box * scale
box[:, 0] = box[:, 0] + padding[1][0]
box[:, 1] = box[:, 1] + padding[0][0]
box[:, 2] = box[:, 2] + padding[1][1]
box[:, 3] = box[:, 3] + padding[0][1]
bb = []
for ix, x in enumerate(box):
bb.append(ia.BoundingBox(x[0], x[1], x[2], x[3], labels[ix]))
bbs = ia.BoundingBoxesOnImage(bb, shape=self.image_size)
seq_det = self.img_aug.to_deterministic()
image_aug = seq_det.augment_images([ig])[0]
bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
box = []
labels = []
for i in range(len(bbs.bounding_boxes)):
after = bbs_aug.bounding_boxes[i]
box.append([after.x1, after.y1, after.x2, after.y2])
labels.append(after.label)
box = np.asarray(box)
box = box / np.asarray([
self.image_size[1], self.image_size[0], self.image_size[1],
self.image_size[0]
])
box = np.clip(box, 0, 1)
return image_aug, box, labels
def pull_item(self, index):
image_path = self.images[index]
json_path = image_path.replace('.jpg', '.json')
bundry = json.loads(open(json_path).read().encode('utf8'))
img = cv2.imread(image_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
labels = []
box = []
for x in bundry['shapes']:
pt = x['points']
labels.append(self.labels_index[x['label']])
box.append([pt[0][0],pt[0][1],pt[2][0],pt[2][1]])
box = np.asarray(box)
if False:
ig, window, scale, padding, crop = utils.resize_image_fixed_size(img, self.image_size)
if len(labels)==0:
return ig, box, labels
box = box * scale
box[:, 0] = box[:, 0] + padding[1][0]
box[:, 1] = box[:, 1] + padding[0][0]
box[:, 2] = box[:, 2] + padding[1][1]
box[:, 3] = box[:, 3] + padding[0][1]
else:
ig, box, labels = utils.crop_image_with_box(img, self.image_size, box, labels)
bb = []
for ix, x in enumerate(box):
bb.append(ia.BoundingBox(x[0],x[1],x[2],x[3],labels[ix]))
bbs = ia.BoundingBoxesOnImage(bb, shape=self.image_size)
seq_det = self.img_aug.to_deterministic()
image_aug = seq_det.augment_images([ig])[0]
bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
box = []
labels = []
for i in range(len(bbs.bounding_boxes)):
after = bbs_aug.bounding_boxes[i]
box.append([after.x1, after.y1, after.x2, after.y2])
labels.append(after.label)
box = np.asarray(box)
box = box / np.asarray([self.image_size[1], self.image_size[0], self.image_size[1], self.image_size[0]])
box = np.clip(box, 0,1)
return image_aug, box, labels
def run(json_pth, save_pth, image_pth, check_dr):
js = json.loads(open(json_pth).read())
lbses = js['label_ix']
new_lbs = dict()
for x in lbses:
new_lbs[lbses[x]] = x
print(new_lbs)
base_dr = image_pth
# base_dr = 'D:/deep_learn_data/luntai/crop/crop'
model = resnet18(num_classes=1000, pretrained=None)
model.fc = nn.Linear(512, 1)
model.load_state_dict(torch.load(check_dr), strict=True)
model.cuda()
model.eval()
for k1 in glob.glob(os.path.join(base_dr, '*.*')):
imag = Image.open(k1)
ig = np.asarray(imag)
img, _, _, _, _ = resize_image_fixed_size(ig, [512, 512])
ig = Image.fromarray(img)
ig = trans(ig)
ig = ig.unsqueeze(0)
ig = torch.autograd.Variable(ig.cuda())
output = model(ig)
output = output.squeeze()
ot = F.sigmoid(output)
print(ot)
pred = ot.ge(0.5).float()
pred_lb = pred.cpu().numpy()
pred_lb = int(pred_lb)
pred_label = new_lbs[pred_lb]
name = k.replace('\\', '/').split('/')[-1]
new_path = os.path.join(save_pth, pred_label)
if not os.path.exists(new_path):
os.makedirs(new_path)
shutil.copy(k1, new_path)
def run(json_pth, save_pth, image_pth, check_dr):
js = json.loads(open(json_pth).read())
lbses = js['label_ix']
new_lbs = dict()
for x in lbses:
new_lbs[lbses[x]] = x
print(new_lbs)
base_dr = image_pth
# base_dr = 'D:/deep_learn_data/luntai/crop/crop'
model = se_resnext50_32x4d(num_classes=1000, pretrained=None)
model.fc1 = nn.Linear(2048, 2)
model.fc2 = nn.Linear(2, len(new_lbs))
model.load_state_dict(torch.load(check_dr), strict=False)
model.cuda()
model.eval()
for k1 in glob.glob(os.path.join(base_dr, '*.*')):
imag = Image.open(k1)
ig = np.asarray(imag)
img, _, _, _, _ = resize_image_fixed_size(ig, [512, 512])
ig = Image.fromarray(img)
ig = trans(ig)
ig = ig.unsqueeze(0)
ig = torch.autograd.Variable(ig.cuda())
fc1, output = model(ig)
output = F.softmax(output, dim=1)
pred = output.data.squeeze(dim=0).cpu().numpy()
pred_lb = np.argmax(pred)
sc = np.max(pred)
pred_label = new_lbs[pred_lb]
name = k.replace('\\', '/').split('/')[-1]
new_path = os.path.join(save_pth, pred_label)
if not os.path.exists(new_path):
os.makedirs(new_path)
shutil.copy(k1, new_path)
def pull_item(self, idx):
image_path = self.data[idx]
js_path = self.data[idx].replace('.jpg', '.json')
js_data = json.loads(open(js_path).read())
label_ix = []
box = []
ig_data = io.imread(image_path)
for b in js_data:
category_name = b['label']
if self.class_mapix.get(category_name, None) is not None:
bound = np.asarray(b['points'])
minx, miny, maxx, maxy = min(bound[:, 0]), min(
bound[:, 1]), max(bound[:, 0]), max(bound[:, 1]),
box.append([minx, miny, maxx, maxy])
label_ix.append(self.class_mapix[category_name])
box = np.asarray(box)
if box.shape[0] == 0:
return None, None, None
if self.is_crop:
ig, box, label_ix = utils.crop_image_with_box(
ig_data, self.image_size, box, label_ix)
else:
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
ig_data, self.image_size)
box = box * scale
box[:, 0] = box[:, 0] + padding[1][0]
box[:, 1] = box[:, 1] + padding[0][0]
box[:, 2] = box[:, 2] + padding[1][1]
box[:, 3] = box[:, 3] + padding[0][1]
box = box / np.asarray([
self.image_size[1], self.image_size[0], self.image_size[1],
self.image_size[0]
])
return ig, box, label_ix
def pull_item(self, index):
image_path = self.images[index]
json_path = image_path.replace('.png', '.json')
bundry = json.loads(open(json_path).read().encode('utf8'))
ig = cv2.imread(image_path)
ig = cv2.cvtColor(ig, cv2.COLOR_BGR2RGB)
ig = aug_utils.pytorch_aug_color(ig)
shape = ig.shape[0:2]
total = len(bundry)
msk = np.zeros((shape[0], shape[1], total), dtype=np.uint8)
ids = []
orchard = np.zeros(shape=(shape[0], shape[1], 3), dtype=np.uint8)
msk = []
for idx, b in enumerate(bundry):
if b['correction_type'] == 'tree':
mask = np.zeros(shape=(shape[0], shape[1], 3), dtype=np.uint8)
pts = []
for p in b['boundary']:
pts.append([int(p['x']), int(p['y'])])
pts = np.array(pts, np.int32)
cv2.fillPoly(mask, [pts], color=(255, 255, 255))
msk.append(mask[:, :, 0:1])
ids.append(0)
elif b['correction_type'] == 'orchard':
mask = np.zeros(shape=(shape[0], shape[1], 3), dtype=np.uint8)
pts = []
for p in b['boundary']:
pts.append([int(p['x']), int(p['y'])])
pts = np.array(pts, np.int32)
cv2.fillPoly(mask, [pts], color=(255, 255, 255))
orchard += mask
ig = ig * (orchard / 255)
if len(msk) > 1:
msk = np.concatenate(msk, axis=2)
elif len(msk) == 1:
msk = msk[0]
else:
return None
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
ig, self.image_size)
msk = utils.resize_mask(msk, scale, padding, crop)
if random.randint(0, 1) != 1:
ag = self.aug.to_deterministic()
ig = ag.augment_image(ig)
msk = ag.augment_image(msk)
box = utils.extract_bboxes(msk)
ids = np.asarray(ids)
mj = (box[:, 3] - box[:, 1]) * (box[:, 2] - box[:, 0])
mk = np.where(mj > self.image_size[0] * self.image_size[1] / 32 / 32)
box = box[mk]
ids = ids[mk]
mj = (box[:, 3] - box[:, 1]) / (box[:, 2] - box[:, 0])
mk = np.where(mj > 0.25)
box = box[mk]
ids = ids[mk]
mj = (box[:, 3] - box[:, 1]) / (box[:, 2] - box[:, 0])
mk = np.where(mj < 4)
box = box[mk]
ids = ids[mk]
box = box / np.asarray([
self.image_size[0], self.image_size[1], self.image_size[0],
self.image_size[1]
])
return ig, box, ids
def pull_item(self, index):
image_path = self.images[index]
json_path = image_path.replace('.png', '.json')
bundry = json.loads(open(json_path).read().encode('utf8'))
ig = cv2.imread(image_path)
ig = cv2.cvtColor(ig, cv2.COLOR_BGR2RGB)
shape = ig.shape[0:2]
orchard = []
msk = []
ids = []
for idx, b in enumerate(bundry):
if b['correction_type'] == 'tree':
mask = np.zeros(shape=(shape[0], shape[1], 3))
pts = []
for p in b['boundary']:
pts.append([int(p['x']), int(p['y'])])
pts = np.array(pts, np.int32)
cv2.fillPoly(mask, [pts], color=(255, 255, 255))
msk.append(mask[:, :, 0:1])
ids.append(0)
elif b['correction_type'] == 'orchard':
mask = np.zeros(shape=(shape[0], shape[1], 3))
pts = []
for p in b['boundary']:
pts.append([int(p['x']), int(p['y'])])
pts = np.array(pts, np.int32)
cv2.fillPoly(mask, [pts], color=(255, 255, 255))
orchard.append(mask[:, :, 0:1])
if len(msk) > 1:
msk = np.concatenate(msk, axis=2)
elif len(msk) == 1:
msk = msk[0]
else:
return None
if len(orchard) > 1:
orchard = np.concatenate(orchard, axis=2)
elif len(orchard) == 1:
orchard = orchard[0]
else:
return None
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
ig, self.image_size)
msk = utils.resize_mask(msk, scale, padding, crop)
orchard = utils.resize_mask(orchard, scale, padding, crop)
if random.randint(0, 1) == 1:
ag = self.aug.to_deterministic()
ig = ag.augment_image(ig)
msk = ag.augment_image(msk)
orchard = ag.augment_image(orchard)
box = utils.extract_bboxes(msk)
ids = np.asarray(ids)
mask = np.transpose(msk, [2, 0, 1])
mj = (box[:, 3] - box[:, 1]) * (box[:, 2] - box[:, 0])
mk = np.where(mj < self.image_size[0] * self.image_size[1] / 3 / 3)
box = box[mk]
ids = ids[mk]
mask = mask[mk]
mj = (box[:, 3] - box[:, 1]) * (box[:, 2] - box[:, 0])
mk = np.where(mj > self.image_size[0] * self.image_size[1] / 32 / 32)
box = box[mk]
ids = ids[mk]
mask = mask[mk]
mj = (box[:, 3] - box[:, 1]) / (box[:, 2] - box[:, 0])
mk = np.where(mj > 0.3)
box = box[mk]
ids = ids[mk]
mask = mask[mk]
mj = (box[:, 3] - box[:, 1]) / (box[:, 2] - box[:, 0])
mk = np.where(mj < 3)
box = box[mk]
ids = ids[mk]
mask = mask[mk]
orchard = np.sum(orchard, axis=2)
mask = np.transpose(mask, [1, 2, 0])
#mask = utils.minimize_mask(box, mask, mini_shape=(28, 28))
mask = np.sum(mask, 2)
mask[np.where(mask > 255)] = 255
box = box / np.asarray([
self.image_size[0], self.image_size[1], self.image_size[0],
self.image_size[1]
])
#ig = ig*(orchard/255).astype(np.uint8)
return ig, box, ids, mask
def pull_item(self, index):
img_id = self.ids[index]
boxes = []
labels = []
target = ET.parse(self._annopath % img_id).getroot()
for obj in target.iter('object'):
difficult = int(obj.find('difficult').text) == 1
if False and difficult:
continue
name = obj.find('name').text.lower().strip()
bbox = obj.find('bndbox')
pts = ['xmin', 'ymin', 'xmax', 'ymax']
bndbox = []
for i, pt in enumerate(pts):
cur_pt = int(bbox.find(pt).text) - 1
bndbox.append(cur_pt)
if self.class_mapix.get(name, None) is not None:
label_idx = self.class_mapix[name]
labels.append(label_idx)
#labels.append(0)
boxes.append(bndbox)
if len(boxes) == 0:
return None, None, None
img = io.imread(self._imgpath % img_id)
box = np.asarray(boxes)
if True:
ig, window, scale, padding, crop = utils.resize_image_fixed_size(
img, self.image_size)
if len(labels) == 0:
return ig, box, labels
box = box * scale
box[:, 0] = box[:, 0] + padding[1][0]
box[:, 1] = box[:, 1] + padding[0][0]
box[:, 2] = box[:, 2] + padding[1][1]
box[:, 3] = box[:, 3] + padding[0][1]
else:
ig, box, labels = utils.crop_image_with_box(
img, self.image_size, box, labels)
bb = []
for ix, x in enumerate(box):
bb.append(ia.BoundingBox(x[0], x[1], x[2], x[3], labels[ix]))
bbs = ia.BoundingBoxesOnImage(bb, shape=self.image_size)
seq_det = self.img_aug.to_deterministic()
image_aug = seq_det.augment_images([ig])[0]
bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
box = []
labels = []
for i in range(len(bbs.bounding_boxes)):
after = bbs_aug.bounding_boxes[i]
box.append([after.x1, after.y1, after.x2, after.y2])
labels.append(after.label)
box = np.asarray(box)
box = box / np.asarray([
self.image_size[1], self.image_size[0], self.image_size[1],
self.image_size[0]
])
box = np.clip(box, 0, 1)
return image_aug, box, labels
