First of all, thank you for providing good open source.
In the process of trying to combining the tiled image in the inference process, it faced difficulties.
data info
original image average w, h = 10000, 8500
average bbox w, h = 188, 166
quantile information of percentage area
0.10 0.014795
0.20 0.022676
0.30 0.028893
0.40 0.031821
0.50 0.038901
0.60 0.043673
0.70 0.051553
0.80 0.066869
0.90 0.101932
0.99 0.317536
99% of the objects occupied less than 0.3% of the original image, and image tiling was performed in the preprocessing step
Tiling was overlapped and duplicated annotation information was deleted.
# tiler code
class SplitImage():
def __init__(self):
self.images_df = []
self.annos_df = []
self.new_image_id = 1
self.new_img_list = []
self.new_anno_id = 0
self.new_anno_list = []
self.slice_bboxes = []
def split(self, im, rows, cols, image_path, img_dict, overlap_height_ratio: float=0.2, overlap_width_ratio: float=0.2):
save_img_dir = r'./data'
os.makedirs(save_img_dir, exist_ok=True)
im_width, im_height = im.size
print('######## image size: ', im.size)
slice_width = int(im_width / rows)
slice_height = int(im_height / cols)
"""
Given the height and width of an image, calculates how to divide the image into
overlapping slices according to the height and width provided. These slices are returned
as bounding boxes in xyxy format
"""
check_list = []
n = 0
y_max = y_min = 0
y_overlap = int(overlap_height_ratio * slice_height)
x_overlap = int(overlap_width_ratio * slice_width)
while y_max < im_height:
x_min = x_max = 0
y_max = y_min + slice_height
while x_max < im_width:
x_max = x_min + slice_width
n += 1
if y_max > im_height or x_max > im_width:
xmax = min(im_width, x_max)
ymax = min(im_height, y_max)
xmin = max(0, xmax - slice_width)
ymin = max(0, ymax - slice_height)
slice_bbox = [xmin, ymin, xmax, ymax]
print('____slice_bbox:', slice_bbox)
new_img_dict = {}
outp = im.crop(slice_bbox)
name, ext = os.path.splitext(os.path.basename(image_path))
new_name = name + "_" + str(n) + ext
outp_path = os.path.join(save_img_dir, new_name)
image_yn = True # annotations 정보 없는 이미지도 저장
# new annotation
for anno_dict in self.annos_df[self.annos_df['image_id'] == img_dict["id"]].to_dict("records"):
new_anno_dict = {}
coord = anno_dict['bbox']
coord_tp = [coord[0], coord[1], coord[0] + coord[2], coord[1] + coord[3]]
if coord_tp[0] >= slice_bbox[0] and coord_tp[2] <= slice_bbox[2]:
if coord_tp[1] >= slice_bbox[1] and coord_tp[3] <= slice_bbox[3]:
# image_yn = True # image_yn이 False로 설정 되어있을 때 사용
if [coord[0], coord[1], coord[2], coord[3]] in check_list:
pass
else:
self.new_anno_id += 1
new_anno_dict["id"] = self.new_anno_id
new_anno_dict['image_id'] = self.new_image_id
new_anno_dict['category_id'] = anno_dict['category_id']
new_anno_dict['category_name'] = anno_dict['category_name']
new_anno_dict['class_name'] = anno_dict['class_name']
new_anno_dict['bbox'] = [round(coord[0]-slice_bbox[0], 1), round(coord[1]-slice_bbox[1], 1), coord[2], coord[3]]
new_anno_dict['area'] = anno_dict['area']
new_anno_dict['iscrowd'] = 0
check_list.append([coord[0], coord[1], coord[2], coord[3]])
self.new_anno_list.append(new_anno_dict)
# new image
if image_yn:
new_img_dict["id"] = self.new_image_id
new_img_dict["file_name"] = new_name
new_img_dict["width"] = outp.size[0]
new_img_dict["height"] = outp.size[1]
new_img_dict["slices"] = slice_bbox
new_img_dict["ori_width"] = im_width
new_img_dict["ori_height"] = im_height
new_img_dict["ori_file_name"] = os.path.basename(image_path)
self.new_img_list.append(new_img_dict)
self.new_image_id += 1
print("Exporting image tile: " + outp_path)
outp.save(outp_path)
else:
slice_bbox = [x_min, y_min, x_max, y_max]
print('slice_bbox:', slice_bbox)
new_img_dict = {}
outp = im.crop(slice_bbox)
name, ext = os.path.splitext(os.path.basename(image_path))
new_name = name + "_" + str(n) + ext
outp_path = os.path.join(save_img_dir, new_name)
image_yn = True # annotations 정보 없는 이미지도 저장
# new annotation
for anno_dict in self.annos_df[self.annos_df['image_id'] == img_dict["id"]].to_dict("records"):
new_anno_dict = {}
coord = anno_dict['bbox']
coord_tp = [coord[0], coord[1], coord[0] + coord[2], coord[1] + coord[3]]
if coord_tp[0] >= slice_bbox[0] and coord_tp[2] <= slice_bbox[2]:
if coord_tp[1] >= slice_bbox[1] and coord_tp[3] <= slice_bbox[3]:
# image_yn = True
if [coord[0], coord[1], coord[2], coord[3]] in check_list:
pass
else:
self.new_anno_id += 1
new_anno_dict["id"] = self.new_anno_id
new_anno_dict['image_id'] = self.new_image_id
new_anno_dict['category_id'] = anno_dict['category_id']
new_anno_dict['category_name'] = anno_dict['category_name']
new_anno_dict['class_name'] = anno_dict['class_name']
new_anno_dict['bbox'] = [round(coord[0]-slice_bbox[0], 1), round(coord[1]-slice_bbox[1], 1), coord[2], coord[3]]
new_anno_dict['area'] = anno_dict['area']
new_anno_dict['iscrowd'] = 0
check_list.append([coord[0], coord[1], coord[2], coord[3]])
self.new_anno_list.append(new_anno_dict)
# new image
if image_yn:
new_img_dict["id"] = self.new_image_id
new_img_dict["file_name"] = new_name
new_img_dict["width"] = outp.size[0]
new_img_dict["height"] = outp.size[1]
new_img_dict["slices"] = slice_bbox
new_img_dict["ori_width"] = im_width
new_img_dict["ori_height"] = im_height
new_img_dict["ori_file_name"] = os.path.basename(image_path)
self.new_img_list.append(new_img_dict)
self.new_image_id += 1
print("Exporting image tile: " + outp_path)
outp.save(outp_path)
x_min = x_max - x_overlap
y_min = y_max - y_overlap
tiled json info [test data set]
tiled image directory [ test dir]
I want to reason with these images and then combine them again. But I just can't think of what to do.
At this time, when combining the separated images back to the original image, I want to delete the duplicated bbox.
I want to know if there's a way to apply it in the inference process. <<< in the eval_coco.py
First of all, thank you for providing good open source.
In the process of trying to combining the tiled image in the inference process, it faced difficulties.
original image average w, h = 10000, 8500 average bbox w, h = 188, 166
quantile information of percentage area
99% of the objects occupied less than 0.3% of the original image, and image tiling was performed in the preprocessing step Tiling was overlapped and duplicated annotation information was deleted.
I want to reason with these images and then combine them again. But I just can't think of what to do. At this time, when combining the separated images back to the original image, I want to delete the duplicated bbox. I want to know if there's a way to apply it in the inference process. <<< in the eval_coco.py