def measure_occlusion(bbox,
trackedObjects,
cropSize=CROP_SIZE,
cropPad=CROP_PAD):
image = np.zeros((int(cropSize), int(cropSize)), dtype=np.bool)
fullBBox = scale_bbox(bbox, cropPad)
fullBBoxXYWH = xyxy_to_xywh(fullBBox)
fullBBoxXYWH[[2, 3]] = np.maximum(fullBBoxXYWH[[2, 3]], 1)
# Now do all objects
for obj in trackedObjects:
boxPos = obj.get_bbox()
boxPosXYWH = xyxy_to_xywh(boxPos)
if IOU(boxPos, fullBBox) < 0.001:
continue
cropCoords = np.zeros(4)
cropCoords = np.clip(boxPos - fullBBox[[0, 1, 0, 1]], 0,
fullBBoxXYWH[[2, 3, 2, 3]])
cropCoords *= cropSize * 1.0 / fullBBoxXYWH[[2, 3, 2, 3]]
cropCoords = np.clip(np.round(cropCoords), 0, cropSize).astype(int)
if (cropCoords[2] - cropCoords[0] < 1
or cropCoords[3] - cropCoords[1] < 1):
cropCoords[[0, 1]] = np.clip(cropCoords[[0, 1]] - 1, 0,
cropSize).astype(int)
cropCoords[[2, 3]] = np.clip(cropCoords[[2, 3]] + 1, 0,
cropSize).astype(int)
image[cropCoords[1]:cropCoords[3], cropCoords[0]:cropCoords[2]] = True
return np.count_nonzero(image) * 1.0 / image.size
def rectangle_guessing_loss(y_true, y_pred): y_true_np = K.eval(y_true) y_pred_np = K.eval(y_pred) y_true_iter = np.nditer(y_true_np, order='C') y_pred_iter = np.nditer(y_pred_np, order='C') loss = 0.0 count = 0 for y1, y2 in zip(y_true_iter, y_pred_iter): loss.append(IOU(y1, y2)) count += 1 return (loss / count)
def responsible_box(self, idx, label, predictions, sx, sy):
max_iou = 0.0
responsible = 0
for b in range(self.B):
boxA, boxB = self.get_real_boxes(label[5 * idx + 1:5 * idx + 5],
predictions[5 * b + 1:5 * b + 5],
sx, sy)
iou = IOU(boxA, boxB)
if max_iou < iou:
max_iou = iou
responsible = b
return b, max_iou
def _kmeans(self, X):
k = self.NAnchors
print('kmeans with k = {0}'.format(k))
x_num = X.shape[0]
iterations = 0
self.dataset['train'].append(ell)
prev_assignments = np.ones(k)*(-1)
iteration = 0
old_distances = np.zeros((x_num, k))
indices = [random.randrange(x_num) for i in range(k)]
centroids = X[indices]
anchor_dim = X.shape[1]
while True:
distances = []
iteration += 1
for i in range(x_num):
d = 1 - IOU(ann_dims[i], centroids)
distances.append(d)
distances = np.array(distances)
print("iteration {}: dists = {}".format(iteration, np.sum(np.abs(old_distances-distances))))
#assign samples to centroids
assignments = np.argmin(distances,axis=1)
if (assignments == prev_assignments).all() :
return centroids
#calculate new centroids
centroid_sums=np.zeros((k, anchor_dim), np.float)
for i in range(x_num):
centroid_sums[assignments[i]]+=X[i]
for j in range(k):
centroids[j] = centroid_sums[j]/(np.sum(assignments==j) + 1e-6)
prev_assignments = assignments.copy()
old_distances = distances.copy()
def remove_bb(boxes):
""" removes redundant bbs from list which have great iou with other bbs in list.
"""
for i in range(len(boxes)):
boxA = boxes[i]
if boxA is None:
continue
boxA_rect = calc_rect(boxA)
for j in range( len(boxes)):
boxB = boxes[j]
if boxB is None:
continue
if np.array_equal(boxA, boxB):
continue
boxB_rect = calc_rect(boxB)
iou = IOU(boxA_rect, boxB_rect)
if iou>0.3:
boxes[j] =None
#filter out None elements from list
boxes_final = list(filter(None, boxes))
return boxes_final
dscores = detection_result["instances"].scores.cpu().detach(
).numpy()
center_pos = tracker.tracker.center_pos
# print(tracker.tracker.center_pos, tracker.tracker.size)
# print(dboxes, dscores)
# print(x_crop.shape)
all_outputs = tracker.track(frame, x_crop, scale_z,
search_instance_size)
# cv2.imwrite("test.jpg", frame)
tboxes, tscores = all_outputs['bbox'], all_outputs['best_score']
# print(tboxes, tscores)
for idx, dbox in enumerate(dboxes):
for tbox in tboxes:
if IOU(dbox, tbox) > 0.8:
dscores[idx] = 1
# print(dscores)
# Windows penalty for detection results
input_size = int(search_instance_size / scale_z)
hanning = np.hanning(input_size)
window = np.outer(hanning, hanning)
idx = 0
for idx, dbox in enumerate(dboxes):
x = int((dbox[0] + dbox[2]) / 2)
y = int((dbox[1] + dbox[3]) / 2)
# print(x, y, idx)
dscores[idx] = dscores[idx] * (
print('start testing {} samples...'.format(num_test_samples))
for ti in range(num_test_samples // batch_size):
x_batch, y_batch = next(test_generator)
# tensorflow wants a different tensor order
feed_dict = {
img: x_batch,
label: y_batch,
}
loss, pred_logits = sess.run([cross_entropy_loss, pred],
feed_dict=feed_dict)
pred_map_batch = np.argmax(pred_logits, axis=3)
x_batch, y_batch = next(train_generator)
feed_dict = {img: x_batch, label: y_batch}
_, loss, summary, lr, pred_logits = sess.run([
train_step, cross_entropy_loss, summary_merged, learning_rate, pred
],
feed_dict=feed_dict)
global_step.assign(it).eval()
train_writer.add_summary(summary, it)
score = IOU(1 / (1 + np.exp(-pred_logits[0])), y_batch[0])
if it % 20 == 0:
try:
print('[iter {}, epoch {}]: lr={} loss={}, IOU={}'.format(
it,
float(it) / steps_per_epoch, lr, loss, score))
except:
pass
idx += 1
height, width, channel = img.shape
neg_num = 0
#先采样一定数量neg图片
while neg_num < 50:
#随机选取截取图像大小
size = npr.randint(12, min(width, height) / 2)
#随机选取左上坐标
nx = npr.randint(0, width - size)
ny = npr.randint(0, height - size)
#截取box
crop_box = np.array([nx, ny, nx + size, ny + size])
#计算iou值
Iou = IOU(crop_box, boxes)
#截取图片并resize成12x12大小
cropped_im = img[ny:ny + size, nx:nx + size, :]
resized_im = cv2.resize(cropped_im, (12, 12),
interpolation=cv2.INTER_LINEAR)
#iou值小于0.3判定为neg图像
if np.max(Iou) < 0.3:
save_file = os.path.join(neg_save_dir, '%s.jpg' % n_idx)
f2.write(neg_save_dir + '/%s.jpg' % n_idx + ' 0\n')
cv2.imwrite(save_file, resized_im)
n_idx += 1
neg_num += 1
for box in boxes:
#左上右下坐标
def box_head_evaluation(self, nms_boxes: list, nms_scores: list, nms_labels: list,
gt_boxes: list, gt_labels: list,
iou_thresh: float = 0.5) -> tuple:
"""
Constructs matches list & scores list for every class
Input
-----
nms_boxes: list:len(bz){(post_NMS_boxes_per_image,4)} ([x1,y1,x2,y2] format)
nms_scores: list:len(bz){(post_NMS_boxes_per_image)} ( the score for the top class for
the regressed box)
nms_labels: list:len(bz){(post_NMS_boxes_per_image)} (top class of each regressed box)
gt_bboxes: list: len(bz){(n_obj, 4)}([x1, y1, x2, y2] format)
gt_labels: list: len(bz) {(n_obj)}
Output
-----
matches: (bz, post_NMS_boxes_per_image, 3)
scores: (bz, post_NMS_boxes_per_image, 3)
num_true: (3, )
num_positives: (3, )
"""
matches = []
scores = []
num_trues = torch.zeros(1, 3)
num_positives = torch.zeros(1, 3)
batch_size = len(nms_labels)
for bz in range(batch_size):
match = torch.zeros(nms_labels[bz].shape[0], 3)
score = torch.zeros(match.shape)
num_true = torch.zeros(1, 3)
num_positive = torch.zeros(num_true.shape)
# calculate trues
for obj_label in gt_labels[bz]:
if obj_label > 0:
num_true[0, obj_label.type(torch.long) - 1] += 1
# calculate positives
for obj_label in nms_labels[bz]:
num_positive[0, obj_label.type(torch.long)] += 1
for i_pred, class_pred in enumerate(nms_labels[bz]):
class_pred_scalar = class_pred.item()
if class_pred_scalar + 1 in gt_labels[bz]:
# retrieve class gt label
i_gt_list = (gt_labels[bz] == class_pred + 1).nonzero(as_tuple=False).squeeze(0)
for i_gt in i_gt_list:
# retrieve masks
box_pred = nms_boxes[bz][i_pred]
box_gt = gt_boxes[bz][i_gt.item()]
# compute IOU
iou = IOU(box_pred, box_gt, mode='corner')
if iou > iou_thresh:
match[i_pred, class_pred] = 1
# no matter how we always store the bbox scores
score[i_pred, class_pred] = nms_scores[bz][i_pred]
matches.append(match)
scores.append(score)
num_trues = torch.cat([num_trues, num_true], dim=0)
num_positives = torch.cat([num_positives, num_positive], dim=0)
return torch.cat(matches), torch.cat(scores), \
torch.sum(num_trues, dim=0), torch.sum(num_positives, dim=0)
def main(size):
'''用于处理带有landmark的数据'''
# size=args.input_size
#是否对图像变换
argument = True
if size == 12:
net = 'PNet'
elif size == 24:
net = 'RNet'
elif size == 48:
net = 'ONet'
image_id = 0
#数据输出路径
OUTPUT = os.path.join(data_dir, str(size))
if not os.path.exists(OUTPUT):
os.mkdir(OUTPUT)
#图片处理后输出路径
dstdir = os.path.join(OUTPUT, 'train_%s_landmark_aug' % (net))
if not os.path.exists(dstdir):
os.mkdir(dstdir)
#label记录txt
ftxt = os.path.join(data_dir, 'trainImageList.txt')
#记录label的txt
f = open(os.path.join(OUTPUT, 'landmark_%d_aug.txt' % (size)), 'w')
#获取图像路径，box，关键点
data = getDataFromTxt(ftxt, data_dir)
idx = 0
for (imgPath, box, landmarkGt) in tqdm(data):
#存储人脸图片和关键点
F_imgs = []
F_landmarks = []
img = cv2.imread(imgPath)
img_h, img_w, img_c = img.shape
gt_box = np.array([box.left, box.top, box.right, box.bottom])
#人脸图片
f_face = img[box.top:box.bottom + 1, box.left:box.right + 1]
#resize成网络输入大小
f_face = cv2.resize(f_face, (size, size))
landmark = np.zeros((5, 2))
for index, one in enumerate(landmarkGt):
#关键点相对于左上坐标偏移量并归一化
rv = ((one[0] - gt_box[0]) / (gt_box[2] - gt_box[0]),
(one[1] - gt_box[1]) / (gt_box[3] - gt_box[1]))
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
if argument:
#对图像变换
idx = idx + 1
x1, y1, x2, y2 = gt_box
gt_w = x2 - x1 + 1
gt_h = y2 - y1 + 1
#除去过小图像
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
for i in range(10):
#随机裁剪图像大小
box_size = npr.randint(int(min(gt_w, gt_h) * 0.8),
np.ceil(1.25 * max(gt_w, gt_h)))
#随机左上坐标偏移量
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
#计算左上坐标
nx1 = int(max(x1 + gt_w / 2 - box_size / 2 + delta_x, 0))
ny1 = int(max(y1 + gt_h / 2 - box_size / 2 + delta_y, 0))
nx2 = nx1 + box_size
ny2 = ny1 + box_size
#除去超过边界的
if nx2 > img_w or ny2 > img_h:
continue
#裁剪边框，图片
crop_box = np.array([nx1, ny1, nx2, ny2])
cropped_im = img[ny1:ny2 + 1, nx1:nx2 + 1, :]
resized_im = cv2.resize(cropped_im, (size, size))
iou = IOU(crop_box, np.expand_dims(gt_box, 0))
#只保留pos图像
if iou > 0.65:
F_imgs.append(resized_im)
#关键点相对偏移
for index, one in enumerate(landmarkGt):
rv = ((one[0] - nx1) / box_size,
(one[1] - ny1) / box_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1, 2)
box = BBox([nx1, ny1, nx2, ny2])
#镜像
if random.choice([0, 1]) > 0:
face_flipped, landmark_flipped = flip(
resized_im, landmark_)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#逆时针翻转
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rorated = rotate(
img, box, box.reprojectLandmark(landmark_), 5)
#关键点偏移
landmark_rorated = box.projectLandmark(
landmark_rorated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rorated.reshape(10))
#左右翻转
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rorated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#顺时针翻转
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rorated = rotate(
img, box, box.reprojectLandmark(landmark_), -5)
#关键点偏移
landmark_rorated = box.projectLandmark(
landmark_rorated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rorated.reshape(10))
#左右翻转
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rorated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
for i in range(len(F_imgs)):
#剔除数据偏移量在[0,1]之间
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0:
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(os.path.join(dstdir, '%d.jpg' % (image_id)), F_imgs[i])
landmarks = list(map(str, list(F_landmarks[i])))
f.write(
os.path.join(dstdir, '%d.jpg' % (image_id)) + ' -2 ' +
' '.join(landmarks) + '\n')
image_id += 1
f.close()
return F_imgs, F_landmarks
def compute_map(dataloader, checkpoint_file, device):
# =========================== Pretrained ===============================
# Put the path were you save the given pretrained model
pretrained_path = '../pretrained/checkpoint680.pth'
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
backbone, rpn = pretrained_models_680(pretrained_path)
backbone = backbone.to(device)
rpn = rpn.to(device)
# ========================== Prep ============================
metric_trackers = []
for i in range(3):
metric_trackers.append(MetricTracker(i))
# ========================= Loading Model ==============================
boxHead = BoxHead(Classes=3, P=7, device=device).to(device)
if torch.cuda.is_available():
checkpoint = torch.load(checkpoint_file)
else:
checkpoint = torch.load(checkpoint_file,
map_location=torch.device('cpu'))
print("[INFO] Weight loaded from checkpoint file: {}".format(
checkpoint_file))
boxHead.load_state_dict(checkpoint['model_state_dict'])
boxHead.eval() # set to eval mode
for iter, data in enumerate(tqdm(dataloader), 0):
images = data['images'].to(device)
assert len(images) == 1, "Only support batch_size == 1"
# if (iter == 50):
# break
labels_gt_all = data['labels'][0].to(device)
bbox_gt_all = data["bbox"][0].to(device)
index = data['index'][0]
prob, clas, boxes = run_inference(images, index, backbone, rpn,
boxHead)
for tracker_i in range(3):
# focus on one class
target_class = tracker_i + 1
labels_gt = labels_gt_all[labels_gt_all == target_class]
bbox_gt = bbox_gt_all[labels_gt_all ==
target_class] #n,4 x,y,w,h
clas_pred = clas[clas == target_class] # m
prob_pred = prob[clas == target_class] # m
boxes_pred = boxes[clas == target_class] # m,4 x1,y1,x2,y2
boxes_pred_xywh = torch.zeros_like(boxes_pred,
dtype=boxes_pred.dtype,
device=boxes_pred.device)
boxes_pred_xywh[:, 0] = (boxes_pred[:, 0] + boxes_pred[:, 2]) / 2
boxes_pred_xywh[:, 1] = (boxes_pred[:, 1] + boxes_pred[:, 3]) / 2
boxes_pred_xywh[:, 2] = boxes_pred[:, 2] - boxes_pred[:, 0]
boxes_pred_xywh[:, 3] = boxes_pred[:, 3] - boxes_pred[:, 1]
# determine if it is a match with bbox
N_gt = len(bbox_gt) #n
N_pred = len(clas_pred) #m
tp_indicator = torch.zeros((N_pred, )) #m
match_indice = torch.zeros((N_pred, )) #m
if (N_pred != 0) and (N_gt != 0):
# IOU matrix
iou_mat = torch.zeros((N_pred, N_gt)) # m,n
for x in range(N_pred):
for y in range(N_gt):
iou_mat[x, y] = IOU(
torch.unsqueeze(boxes_pred_xywh[x, :], 0),
torch.unsqueeze(bbox_gt[y, :], 0))
if torch.max(iou_mat[x, :]) >= 0.5:
tp_indicator[x] = 1
index = torch.argmax(iou_mat[x, :])
match_indice[x] = index
metric_trackers[tracker_i].add_match(prob_pred, tp_indicator,
match_indice, N_gt)
# compute map
for i in range(3):
metric_trackers[i].compute_precision_recall()
recall, precision = metric_trackers[i].sorted_pr_curve()
ap = metric_trackers[i].compute_ap()
print("class_id: {}. ap: {}".format(i, ap))
metric_trackers[i].reset()
# cv2.imwrite("test.jpg", np.squeeze(x_crop).transpose((1,2,0)))
dboxes = detection_result["instances"].pred_boxes.tensor.cpu().detach().numpy()
dscores = detection_result["instances"].scores.cpu().detach().numpy()
center_pos = tracker.tracker.center_pos
# print(tracker.tracker.center_pos, tracker.tracker.size)
# print("detection results: ", len(dboxes), dscores)
# print(x_crop.shape)
all_outputs = tracker.track(frame, x_crop, scale_z, search_instance_size)
# cv2.imwrite("test.jpg", frame)
tboxes, tscores = all_outputs['bbox'], all_outputs['best_score']
# print("tracking results: ", tboxes, tscores)
for idx, dbox in enumerate(dboxes):
for tbox in tboxes:
if IOU(dbox, tbox) > 0.8:
dscores[idx] = 1
# print(dscores)
# Windows penalty for detection results
input_size = int(search_instance_size/scale_z)
hanning = np.hanning(input_size)
window = np.outer(hanning, hanning)
idx = 0
for idx, dbox in enumerate(dboxes):
x = int((dbox[0] + dbox[2])/2)
y = int((dbox[1] + dbox[3])/2)
# print(x, y, idx)
dscores[idx] = dscores[idx] * (1 - window_influence) + window[x, y]*window_influence
bboxes = np.array(bbox, dtype=np.float32).reshape(-1, 4)
img = cv2.imread(os.path.join(img_dir, img_path + '.jpg'))
idx += 1
if idx % 100 == 0:
print idx, 'images done!'
height, width, channel = img.shape
neg_num = 0
while neg_num < 50:
size = npr.randint(12, min(width, height) / 2) # 12 or 40
nx = npr.randint(0, width - size)
ny = npr.randint(0, height - size)
crop_box = np.array([nx, ny, nx + size, ny + size])
iou = IOU(crop_box, bboxes)
cropped_img = img[ny:ny + size, nx:nx + size, :]
resized_img = cv2.resize(cropped_img, (12, 12),
interpolation=cv2.INTER_LINEAR)
if np.max(iou) < 0.3:
save_file = os.path.join(neg_save_dir, '%s.jpg' % n_idx)
f2.write('12/negative/%s' % n_idx + ' 0\n')
cv2.imwrite(save_file, resized_img)
n_idx += 1
neg_num += 1
for box in bboxes:
# for each ground truth bbox
x1, y1, x2, y2 = box
w = x2 - x1 + 1
scores = []
tracker = Re3Tracker()
for video in VIDEO_PATH:
tracker.reset()
print('Now in folder: ' + video)
image_path = [video + 'color/' + f for f in os.listdir(video + 'color/')]
image_path.sort()
ground_truth = np.loadtxt(video + 'groundtruth.txt', delimiter=',')
img = cv2.imread(image_path[0])
bbox = to_xyxy(ground_truth[0,:])
bbox = tracker.track(1, img, bbox)
for i in range(1, len(image_path)):
img = cv2.imread(image_path[i])
predicted_bbox = tracker.track(1,img)
score = IOU.IOU(predicted_bbox, to_xyxy(ground_truth[i,:]))
print('score:', score)
scores.append(score)
print('Save scores to file ...')
with open('IOU_scores.txt', 'w') as f:
for s in scores:
f.write("%s\n" % s)
print('File saved.')
def render_patch(bbox,
background,
trackedObjects,
cropSize=CROP_SIZE,
cropPad=CROP_PAD):
bboxXYWH = xyxy_to_xywh(bbox)
image = np.zeros((int(cropSize), int(cropSize), 3), dtype=np.uint8)
fullBBoxXYWH = bboxXYWH.copy()
fullBBoxXYWH[[2, 3]] *= cropPad
fullBBox = xywh_to_xyxy(fullBBoxXYWH)
fullBBoxXYWH = fullBBoxXYWH
fullBBoxXYWH[[2, 3]] = np.maximum(fullBBoxXYWH[[2, 3]], 1)
# First do background
boxPos = np.array([0, 0, background.shape[1], background.shape[0]])
boxPosXYWH = xyxy_to_xywh(boxPos)
cropCoords = np.clip(boxPos - fullBBox[[0, 1, 0, 1]], 0,
fullBBoxXYWH[[2, 3, 2, 3]])
cropCoords *= (cropSize) * 1.0 / fullBBoxXYWH[[2, 3, 2, 3]]
cropCoords = np.clip(np.round(cropCoords), 0, cropSize).astype(int)
textureCrop = np.zeros(4)
textureCrop[0] = int(
max(fullBBox[0] - boxPos[0], 0) * background.shape[1] * 1.0 /
boxPosXYWH[2])
textureCrop[1] = int(
max(fullBBox[1] - boxPos[1], 0) * background.shape[0] * 1.0 /
boxPosXYWH[3])
textureCrop[2] = int(
min((fullBBox[2] - boxPos[0]) * 1.0 / boxPosXYWH[2], 1) *
background.shape[1])
textureCrop[3] = int(
min((fullBBox[3] - boxPos[1]) * 1.0 / boxPosXYWH[3], 1) *
background.shape[0])
if (textureCrop[2] - textureCrop[0] < 1
or textureCrop[3] - textureCrop[1] < 1):
textureCrop = [0, 0, 1, 1]
textureCrop = np.round(textureCrop).astype(int)
textureCrop[[0, 2]] = np.clip(textureCrop[[0, 2]], 0, background.shape[1])
textureCrop[[1, 3]] = np.clip(textureCrop[[1, 3]], 0, background.shape[0])
if cropCoords[3] > cropCoords[1] + 1 and cropCoords[2] > cropCoords[0] + 1:
image[cropCoords[1]:cropCoords[3],
cropCoords[0]:cropCoords[2], :] = (cv2.resize(
background[textureCrop[1]:textureCrop[3],
textureCrop[0]:textureCrop[2], :],
(cropCoords[2] - cropCoords[0],
cropCoords[3] - cropCoords[1])))
# Now do all objects
for obj in trackedObjects:
boxPos = obj.get_bbox()
boxPosXYWH = xyxy_to_xywh(boxPos)
if IOU(boxPos, fullBBox) < 0.001:
continue
cropCoords = np.zeros(4)
cropCoords = np.clip(boxPos - fullBBox[[0, 1, 0, 1]], 0,
fullBBoxXYWH[[2, 3, 2, 3]])
cropCoords *= cropSize * 1.0 / fullBBoxXYWH[[2, 3, 2, 3]]
cropCoords = np.clip(np.round(cropCoords), 0, cropSize).astype(int)
if (cropCoords[2] - cropCoords[0] < 1
or cropCoords[3] - cropCoords[1] < 1):
cropCoords[[0, 1]] = np.clip(cropCoords[[0, 1]] - 1, 0,
cropSize).astype(int)
cropCoords[[2, 3]] = np.clip(cropCoords[[2, 3]] + 1, 0,
cropSize).astype(int)
textureCrop = np.zeros(4, dtype=int)
textureCrop[0] = int(
max(fullBBox[0] - boxPos[0], 0) * obj.texture.shape[1] * 1.0 /
boxPosXYWH[2])
textureCrop[1] = int(
max(fullBBox[1] - boxPos[1], 0) * obj.texture.shape[0] * 1.0 /
boxPosXYWH[3])
textureCrop[2] = int(
min((fullBBox[2] - boxPos[0]) * 1.0 / boxPosXYWH[2], 1) *
obj.texture.shape[1])
textureCrop[3] = int(
min((fullBBox[3] - boxPos[1]) * 1.0 / boxPosXYWH[3], 1) *
obj.texture.shape[0])
if (textureCrop[2] - textureCrop[0] < 1
or textureCrop[3] - textureCrop[1] < 1):
textureCrop = [0, 0, 2, 2]
textureCrop = np.round(textureCrop).astype(int)
textureCrop[[0, 2]] = np.clip(textureCrop[[0, 2]], 0,
obj.texture.shape[1])
textureCrop[[1, 3]] = np.clip(textureCrop[[1, 3]], 0,
obj.texture.shape[0])
# Feathering
currentIm = image[cropCoords[1]:cropCoords[3],
cropCoords[0]:cropCoords[2], :].astype(np.float32)
newIm = cv2.resize(
obj.texture[textureCrop[1]:textureCrop[3],
textureCrop[0]:textureCrop[2], :],
(cropCoords[2] - cropCoords[0], cropCoords[3] - cropCoords[1]),
).astype(np.float32)
if (cropCoords[2] - cropCoords[0] < 1
or cropCoords[3] - cropCoords[1] < 1):
featherWeightOn = 0
else:
featherCrop = np.zeros(4)
featherCrop[0] = int(
max(fullBBox[0] - boxPos[0], 0) *
FEATHER_WEIGHT_ARRAY.shape[1] * 1.0 / boxPosXYWH[2])
featherCrop[1] = int(
max(fullBBox[1] - boxPos[1], 0) *
FEATHER_WEIGHT_ARRAY.shape[0] * 1.0 / boxPosXYWH[3])
featherCrop[2] = int(
min((fullBBox[2] - boxPos[0]) * 1.0 / boxPosXYWH[2], 1) *
FEATHER_WEIGHT_ARRAY.shape[1])
featherCrop[3] = int(
min((fullBBox[3] - boxPos[1]) * 1.0 / boxPosXYWH[3], 1) *
FEATHER_WEIGHT_ARRAY.shape[0])
if (featherCrop[2] - featherCrop[0] < 1
or featherCrop[3] - featherCrop[1] < 1):
featherCrop = [
int(CROP_SIZE / 2 - 1),
int(CROP_SIZE / 2 - 1),
int(CROP_SIZE / 2),
int(CROP_SIZE / 2)
]
featherCrop = np.round(featherCrop).astype(int)
featherCrop[[0, 2]] = np.clip(featherCrop[[0, 2]], 0,
FEATHER_WEIGHT_ARRAY.shape[1])
featherCrop[[1, 3]] = np.clip(featherCrop[[1, 3]], 0,
FEATHER_WEIGHT_ARRAY.shape[0])
featherWeightOn = cv2.resize(
FEATHER_WEIGHT_ARRAY[featherCrop[1]:featherCrop[3],
featherCrop[0]:featherCrop[2], :],
(cropCoords[2] - cropCoords[0],
cropCoords[3] - cropCoords[1])).astype(np.float32) / 255.0
image[cropCoords[1]:cropCoords[3], cropCoords[0]:cropCoords[2], :] = (
(newIm * featherWeightOn + currentIm *
(1 - featherWeightOn)).astype(np.uint8))
return image
