Spaces:
Sleeping
Sleeping
Update utils.py
Browse files
utils.py
CHANGED
@@ -1,49 +1,58 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
import os
|
3 |
import io
|
4 |
import base64
|
5 |
import time
|
6 |
-
from PIL import Image, ImageDraw, ImageFont
|
7 |
-
import json
|
8 |
-
import requests
|
9 |
-
# utility function
|
10 |
-
import os
|
11 |
-
from openai import AzureOpenAI
|
12 |
-
|
13 |
import json
|
14 |
import sys
|
15 |
-
import
|
16 |
-
import
|
|
|
|
|
17 |
import numpy as np
|
18 |
-
|
|
|
19 |
from matplotlib import pyplot as plt
|
|
|
20 |
import easyocr
|
21 |
from paddleocr import PaddleOCR
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
reader = easyocr.Reader(['en'])
|
23 |
paddle_ocr = PaddleOCR(
|
24 |
-
lang='en', # other
|
25 |
use_angle_cls=False,
|
26 |
-
use_gpu=False, # using cuda
|
27 |
show_log=False,
|
28 |
max_batch_size=1024,
|
29 |
use_dilation=True, # improves accuracy
|
30 |
det_db_score_mode='slow', # improves accuracy
|
31 |
-
rec_batch_num=1024
|
32 |
-
|
33 |
-
import base64
|
34 |
-
|
35 |
-
import os
|
36 |
-
import ast
|
37 |
-
import torch
|
38 |
-
from typing import Tuple, List
|
39 |
-
from torchvision.ops import box_convert
|
40 |
-
import re
|
41 |
-
from torchvision.transforms import ToPILImage
|
42 |
-
import supervision as sv
|
43 |
-
import torchvision.transforms as T
|
44 |
|
45 |
|
46 |
def get_caption_model_processor(model_name, model_name_or_path="Salesforce/blip2-opt-2.7b", device=None):
|
|
|
|
|
|
|
|
|
47 |
if not device:
|
48 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
49 |
if model_name == "blip2":
|
@@ -51,44 +60,62 @@ def get_caption_model_processor(model_name, model_name_or_path="Salesforce/blip2
|
|
51 |
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
|
52 |
if device == 'cpu':
|
53 |
model = Blip2ForConditionalGeneration.from_pretrained(
|
54 |
-
|
55 |
-
|
56 |
else:
|
57 |
model = Blip2ForConditionalGeneration.from_pretrained(
|
58 |
-
|
59 |
-
|
60 |
elif model_name == "florence2":
|
61 |
-
from transformers import AutoProcessor, AutoModelForCausalLM
|
62 |
processor = AutoProcessor.from_pretrained("microsoft/Florence-2-base", trust_remote_code=True)
|
63 |
if device == 'cpu':
|
64 |
-
model = AutoModelForCausalLM.from_pretrained(
|
|
|
|
|
65 |
else:
|
66 |
-
model = AutoModelForCausalLM.from_pretrained(
|
|
|
|
|
67 |
return {'model': model.to(device), 'processor': processor}
|
68 |
|
69 |
|
70 |
def get_yolo_model(model_path):
|
|
|
|
|
|
|
71 |
from ultralytics import YOLO
|
72 |
-
# Load the model.
|
73 |
model = YOLO(model_path)
|
74 |
return model
|
75 |
|
76 |
|
77 |
@torch.inference_mode()
|
78 |
def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=None, batch_size=32):
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
to_pil = ToPILImage()
|
82 |
if starting_idx:
|
83 |
non_ocr_boxes = filtered_boxes[starting_idx:]
|
84 |
else:
|
85 |
non_ocr_boxes = filtered_boxes
|
86 |
-
|
87 |
-
for
|
88 |
-
xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
|
89 |
-
ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
|
90 |
cropped_image = image_source[ymin:ymax, xmin:xmax, :]
|
91 |
-
|
92 |
|
93 |
model, processor = caption_model_processor['model'], caption_model_processor['processor']
|
94 |
if not prompt:
|
@@ -99,17 +126,29 @@ def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_
|
|
99 |
|
100 |
generated_texts = []
|
101 |
device = model.device
|
102 |
-
for i in range(0, len(
|
103 |
-
|
104 |
-
batch = croped_pil_image[i:i+batch_size]
|
105 |
if model.device.type == 'cuda':
|
106 |
-
inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt").to(device=device, dtype=torch.float16)
|
107 |
else:
|
108 |
-
inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt").to(device=device)
|
109 |
if 'florence' in model.config.name_or_path:
|
110 |
-
generated_ids = model.generate(
|
|
|
|
|
|
|
|
|
|
|
|
|
111 |
else:
|
112 |
-
generated_ids = model.generate(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
|
114 |
generated_text = [gen.strip() for gen in generated_text]
|
115 |
generated_texts.extend(generated_text)
|
@@ -117,52 +156,56 @@ def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_
|
|
117 |
return generated_texts
|
118 |
|
119 |
|
120 |
-
|
121 |
def get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor):
|
|
|
|
|
|
|
122 |
to_pil = ToPILImage()
|
123 |
if ocr_bbox:
|
124 |
non_ocr_boxes = filtered_boxes[len(ocr_bbox):]
|
125 |
else:
|
126 |
non_ocr_boxes = filtered_boxes
|
127 |
-
|
128 |
-
for
|
129 |
-
xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
|
130 |
-
ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
|
131 |
cropped_image = image_source[ymin:ymax, xmin:xmax, :]
|
132 |
-
|
133 |
|
134 |
model, processor = caption_model_processor['model'], caption_model_processor['processor']
|
135 |
device = model.device
|
136 |
-
messages = [{"role": "user", "content": "<|image_1|>\ndescribe the icon in one sentence"}]
|
137 |
prompt = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
138 |
|
139 |
batch_size = 5 # Number of samples per batch
|
140 |
generated_texts = []
|
141 |
|
142 |
-
for i in range(0, len(
|
143 |
-
images =
|
144 |
image_inputs = [processor.image_processor(x, return_tensors="pt") for x in images]
|
145 |
-
inputs ={'input_ids': [], 'attention_mask': [], 'pixel_values': [], 'image_sizes': []}
|
146 |
texts = [prompt] * len(images)
|
147 |
-
for
|
148 |
-
|
149 |
-
inputs['input_ids'].append(
|
150 |
-
inputs['attention_mask'].append(
|
151 |
-
inputs['pixel_values'].append(
|
152 |
-
inputs['image_sizes'].append(
|
153 |
-
max_len = max(
|
154 |
-
for
|
155 |
-
|
156 |
-
inputs['
|
|
|
|
|
157 |
inputs_cat = {k: torch.concatenate(v).to(device) for k, v in inputs.items()}
|
158 |
|
159 |
-
generation_args = {
|
160 |
-
"max_new_tokens": 25,
|
161 |
-
"temperature": 0.01,
|
162 |
-
"do_sample": False,
|
163 |
-
}
|
164 |
-
generate_ids = model.generate(**inputs_cat, eos_token_id=processor.tokenizer.eos_token_id, **generation_args)
|
165 |
-
#
|
166 |
generate_ids = generate_ids[:, inputs_cat['input_ids'].shape[1]:]
|
167 |
response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
168 |
response = [res.strip('\n').strip() for res in response]
|
@@ -170,7 +213,19 @@ def get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, captio
|
|
170 |
|
171 |
return generated_texts
|
172 |
|
|
|
173 |
def remove_overlap(boxes, iou_threshold, ocr_bbox=None):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
174 |
assert ocr_bbox is None or isinstance(ocr_bbox, List)
|
175 |
|
176 |
def box_area(box):
|
@@ -184,39 +239,30 @@ def remove_overlap(boxes, iou_threshold, ocr_bbox=None):
|
|
184 |
return max(0, x2 - x1) * max(0, y2 - y1)
|
185 |
|
186 |
def IoU(box1, box2):
|
187 |
-
|
188 |
-
union = box_area(box1) + box_area(box2) -
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
else:
|
193 |
-
ratio1, ratio2 = 0, 0
|
194 |
-
return max(intersection / union, ratio1, ratio2)
|
195 |
|
196 |
def is_inside(box1, box2):
|
197 |
-
|
198 |
-
|
199 |
-
ratio1 = intersection / box_area(box1)
|
200 |
-
return ratio1 > 0.95
|
201 |
|
202 |
boxes = boxes.tolist()
|
203 |
filtered_boxes = []
|
204 |
if ocr_bbox:
|
205 |
filtered_boxes.extend(ocr_bbox)
|
206 |
-
# print('ocr_bbox!!!', ocr_bbox)
|
207 |
for i, box1 in enumerate(boxes):
|
208 |
-
# if not any(IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2) for j, box2 in enumerate(boxes) if i != j):
|
209 |
is_valid_box = True
|
210 |
for j, box2 in enumerate(boxes):
|
211 |
-
# keep the smaller box
|
212 |
if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
|
213 |
is_valid_box = False
|
214 |
break
|
215 |
if is_valid_box:
|
216 |
-
# add the following 2 lines to include ocr bbox
|
217 |
if ocr_bbox:
|
218 |
-
#
|
219 |
-
if not any(IoU(box1, box3) > iou_threshold and not is_inside(box1, box3) for
|
220 |
filtered_boxes.append(box1)
|
221 |
else:
|
222 |
filtered_boxes.append(box1)
|
@@ -224,11 +270,17 @@ def remove_overlap(boxes, iou_threshold, ocr_bbox=None):
|
|
224 |
|
225 |
|
226 |
def remove_overlap_new(boxes, iou_threshold, ocr_bbox=None):
|
227 |
-
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
232 |
assert ocr_bbox is None or isinstance(ocr_bbox, List)
|
233 |
|
234 |
def box_area(box):
|
@@ -242,99 +294,91 @@ def remove_overlap_new(boxes, iou_threshold, ocr_bbox=None):
|
|
242 |
return max(0, x2 - x1) * max(0, y2 - y1)
|
243 |
|
244 |
def IoU(box1, box2):
|
245 |
-
|
246 |
-
union = box_area(box1) + box_area(box2) -
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
else:
|
251 |
-
ratio1, ratio2 = 0, 0
|
252 |
-
return max(intersection / union, ratio1, ratio2)
|
253 |
|
254 |
def is_inside(box1, box2):
|
255 |
-
|
256 |
-
|
257 |
-
ratio1 = intersection / box_area(box1)
|
258 |
-
return ratio1 > 0.80
|
259 |
|
260 |
-
# boxes = boxes.tolist()
|
261 |
filtered_boxes = []
|
262 |
if ocr_bbox:
|
263 |
filtered_boxes.extend(ocr_bbox)
|
264 |
-
# print('ocr_bbox!!!', ocr_bbox)
|
265 |
for i, box1_elem in enumerate(boxes):
|
266 |
box1 = box1_elem['bbox']
|
267 |
is_valid_box = True
|
268 |
for j, box2_elem in enumerate(boxes):
|
269 |
-
# keep the smaller box
|
270 |
box2 = box2_elem['bbox']
|
271 |
if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
|
272 |
is_valid_box = False
|
273 |
break
|
274 |
if is_valid_box:
|
275 |
-
# add the following 2 lines to include ocr bbox
|
276 |
if ocr_bbox:
|
277 |
-
# keep yolo boxes + prioritize ocr label
|
278 |
box_added = False
|
279 |
for box3_elem in ocr_bbox:
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
# break
|
292 |
-
elif is_inside(box1, box3): # icon inside ocr
|
293 |
-
box_added = True
|
294 |
-
# try:
|
295 |
-
# filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': None})
|
296 |
-
# filtered_boxes.remove(box3_elem)
|
297 |
-
# except:
|
298 |
-
# continue
|
299 |
-
break
|
300 |
-
else:
|
301 |
continue
|
|
|
|
|
|
|
302 |
if not box_added:
|
303 |
-
filtered_boxes.append({
|
304 |
-
|
|
|
|
|
|
|
|
|
305 |
else:
|
306 |
filtered_boxes.append(box1)
|
307 |
-
return filtered_boxes
|
308 |
|
309 |
|
310 |
def load_image(image_path: str) -> Tuple[np.array, torch.Tensor]:
|
311 |
-
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
|
|
|
|
|
|
|
|
|
|
318 |
image_source = Image.open(image_path).convert("RGB")
|
319 |
image = np.asarray(image_source)
|
320 |
image_transformed, _ = transform(image_source, None)
|
321 |
return image, image_transformed
|
322 |
|
323 |
|
324 |
-
def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor, phrases: List[str],
|
325 |
-
text_padding=5, text_thickness=2, thickness=3) -> np.ndarray:
|
326 |
-
"""
|
327 |
-
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
Returns:
|
337 |
-
|
338 |
"""
|
339 |
h, w, _ = image_source.shape
|
340 |
boxes = boxes * torch.Tensor([w, h, w, h])
|
@@ -344,30 +388,43 @@ def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor
|
|
344 |
|
345 |
labels = [f"{phrase}" for phrase in range(boxes.shape[0])]
|
346 |
|
347 |
-
from
|
348 |
-
box_annotator
|
|
|
|
|
349 |
annotated_frame = image_source.copy()
|
350 |
-
annotated_frame = box_annotator.annotate(scene=annotated_frame, detections=detections, labels=labels, image_size=(w,h))
|
351 |
|
352 |
label_coordinates = {f"{phrase}": v for phrase, v in zip(phrases, xywh)}
|
353 |
return annotated_frame, label_coordinates
|
354 |
|
355 |
|
356 |
def predict(model, image, caption, box_threshold, text_threshold):
|
357 |
-
""" Use huggingface model to replace the original model
|
358 |
"""
|
359 |
-
|
360 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
361 |
|
362 |
inputs = processor(images=image, text=caption, return_tensors="pt").to(device)
|
363 |
with torch.no_grad():
|
364 |
-
outputs =
|
365 |
|
366 |
results = processor.post_process_grounded_object_detection(
|
367 |
outputs,
|
368 |
inputs.input_ids,
|
369 |
-
box_threshold=box_threshold,
|
370 |
-
text_threshold=text_threshold,
|
371 |
target_sizes=[image.size[::-1]]
|
372 |
)[0]
|
373 |
boxes, logits, phrases = results["boxes"], results["scores"], results["labels"]
|
@@ -375,71 +432,106 @@ def predict(model, image, caption, box_threshold, text_threshold):
|
|
375 |
|
376 |
|
377 |
def predict_yolo(model, image_path, box_threshold, imgsz, scale_img, iou_threshold=0.7):
|
378 |
-
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
379 |
kwargs = {
|
380 |
-
'conf': box_threshold, #
|
381 |
-
'iou': iou_threshold, #
|
382 |
'verbose': False
|
383 |
}
|
384 |
-
|
385 |
if scale_img:
|
386 |
kwargs['imgsz'] = imgsz
|
387 |
-
|
388 |
results = model.predict(image_path, **kwargs)
|
389 |
boxes = results[0].boxes.xyxy
|
390 |
conf = results[0].boxes.conf
|
391 |
return boxes, conf, [str(i) for i in range(len(boxes))]
|
392 |
|
393 |
|
394 |
-
def get_som_labeled_img(img_path, model=None, BOX_TRESHOLD
|
395 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
396 |
"""
|
397 |
image_source = Image.open(img_path).convert("RGB")
|
398 |
w, h = image_source.size
|
399 |
if not imgsz:
|
400 |
imgsz = (h, w)
|
401 |
-
#
|
402 |
-
xyxy, logits, phrases = predict_yolo(
|
|
|
|
|
|
|
403 |
xyxy = xyxy / torch.Tensor([w, h, w, h]).to(xyxy.device)
|
404 |
-
|
405 |
phrases = [str(i) for i in range(len(phrases))]
|
406 |
|
407 |
-
#
|
408 |
-
h, w, _ = image_source.shape
|
409 |
if ocr_bbox:
|
410 |
ocr_bbox = torch.tensor(ocr_bbox) / torch.Tensor([w, h, w, h])
|
411 |
-
ocr_bbox=ocr_bbox.tolist()
|
412 |
else:
|
413 |
print('no ocr bbox!!!')
|
414 |
ocr_bbox = None
|
415 |
-
# filtered_boxes = remove_overlap(boxes=xyxy, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox)
|
416 |
-
# starting_idx = len(ocr_bbox)
|
417 |
-
# print('len(filtered_boxes):', len(filtered_boxes), starting_idx)
|
418 |
|
419 |
-
ocr_bbox_elem = [{'type': 'text', 'bbox':box, 'interactivity':False, 'content':txt}
|
420 |
-
|
|
|
|
|
421 |
filtered_boxes = remove_overlap_new(boxes=xyxy_elem, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox_elem)
|
422 |
|
423 |
-
#
|
424 |
filtered_boxes_elem = sorted(filtered_boxes, key=lambda x: x['content'] is None)
|
425 |
-
# get the index of the first 'content': None
|
426 |
starting_idx = next((i for i, box in enumerate(filtered_boxes_elem) if box['content'] is None), -1)
|
427 |
-
|
428 |
|
429 |
-
|
430 |
-
# get parsed icon local semantics
|
431 |
if use_local_semantics:
|
432 |
caption_model = caption_model_processor['model']
|
433 |
-
if 'phi3_v' in caption_model.config.model_type:
|
434 |
-
parsed_content_icon = get_parsed_content_icon_phi3v(
|
435 |
else:
|
436 |
-
parsed_content_icon = get_parsed_content_icon(
|
437 |
ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
|
438 |
icon_start = len(ocr_text)
|
439 |
parsed_content_icon_ls = []
|
440 |
-
#
|
441 |
-
for
|
442 |
-
if box['content'] is None:
|
443 |
box['content'] = parsed_content_icon.pop(0)
|
444 |
for i, txt in enumerate(parsed_content_icon):
|
445 |
parsed_content_icon_ls.append(f"Icon Box ID {str(i+icon_start)}: {txt}")
|
@@ -448,51 +540,72 @@ def get_som_labeled_img(img_path, model=None, BOX_TRESHOLD = 0.01, output_coord_
|
|
448 |
ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
|
449 |
parsed_content_merged = ocr_text
|
450 |
|
451 |
-
|
452 |
-
|
453 |
-
phrases = [i for i in range(len(filtered_boxes))]
|
454 |
|
455 |
-
#
|
456 |
if draw_bbox_config:
|
457 |
-
annotated_frame, label_coordinates = annotate(
|
|
|
|
|
458 |
else:
|
459 |
-
annotated_frame, label_coordinates = annotate(
|
|
|
|
|
|
|
460 |
|
461 |
pil_img = Image.fromarray(annotated_frame)
|
462 |
buffered = io.BytesIO()
|
463 |
pil_img.save(buffered, format="PNG")
|
464 |
encoded_image = base64.b64encode(buffered.getvalue()).decode('ascii')
|
|
|
465 |
if output_coord_in_ratio:
|
466 |
-
|
467 |
-
label_coordinates = {k: [v[0]/w, v[1]/h, v[2]/w, v[3]/h] for k, v in label_coordinates.items()}
|
468 |
assert w == annotated_frame.shape[1] and h == annotated_frame.shape[0]
|
469 |
|
470 |
return encoded_image, label_coordinates, filtered_boxes_elem
|
471 |
|
472 |
|
473 |
def get_xywh(input):
|
474 |
-
|
475 |
-
|
476 |
-
|
|
|
|
|
|
|
|
|
|
|
477 |
|
478 |
def get_xyxy(input):
|
479 |
-
|
480 |
-
|
481 |
-
|
|
|
|
|
|
|
|
|
482 |
|
483 |
def get_xywh_yolo(input):
|
484 |
-
|
485 |
-
|
486 |
-
|
487 |
-
|
|
|
|
|
|
|
488 |
|
489 |
|
490 |
-
def check_ocr_box(image_path, display_img
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
491 |
if use_paddleocr:
|
492 |
-
if easyocr_args is None
|
493 |
-
text_threshold = 0.5
|
494 |
-
else:
|
495 |
-
text_threshold = easyocr_args['text_threshold']
|
496 |
result = paddle_ocr.ocr(image_path, cls=False)[0]
|
497 |
conf = [item[1] for item in result]
|
498 |
coord = [item[0] for item in result if item[1][1] > text_threshold]
|
@@ -501,28 +614,21 @@ def check_ocr_box(image_path, display_img = True, output_bb_format='xywh', goal_
|
|
501 |
if easyocr_args is None:
|
502 |
easyocr_args = {}
|
503 |
result = reader.readtext(image_path, **easyocr_args)
|
504 |
-
# print('goal filtering pred:', result[-5:])
|
505 |
coord = [item[0] for item in result]
|
506 |
text = [item[1] for item in result]
|
507 |
-
|
508 |
if display_img:
|
509 |
opencv_img = cv2.imread(image_path)
|
510 |
opencv_img = cv2.cvtColor(opencv_img, cv2.COLOR_RGB2BGR)
|
511 |
bb = []
|
512 |
for item in coord:
|
513 |
x, y, a, b = get_xywh(item)
|
514 |
-
# print(x, y, a, b)
|
515 |
bb.append((x, y, a, b))
|
516 |
-
cv2.rectangle(opencv_img, (x, y), (x+a, y+b), (0, 255, 0), 2)
|
517 |
-
|
518 |
-
# Display the image
|
519 |
plt.imshow(opencv_img)
|
520 |
else:
|
521 |
if output_bb_format == 'xywh':
|
522 |
bb = [get_xywh(item) for item in coord]
|
523 |
elif output_bb_format == 'xyxy':
|
524 |
bb = [get_xyxy(item) for item in coord]
|
525 |
-
# print('bounding box!!!', bb)
|
526 |
return (text, bb), goal_filtering
|
527 |
-
|
528 |
-
|
|
|
1 |
+
"""
|
2 |
+
utils.py
|
3 |
+
|
4 |
+
This module contains utility functions for:
|
5 |
+
- Loading and processing images
|
6 |
+
- Object detection with YOLO
|
7 |
+
- OCR with EasyOCR / PaddleOCR
|
8 |
+
- Image annotation and bounding box manipulation
|
9 |
+
- Captioning / semantic parsing of detected icons
|
10 |
+
"""
|
11 |
+
|
12 |
import os
|
13 |
import io
|
14 |
import base64
|
15 |
import time
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
import json
|
17 |
import sys
|
18 |
+
import re
|
19 |
+
from typing import Tuple, List
|
20 |
+
|
21 |
+
import torch
|
22 |
import numpy as np
|
23 |
+
import cv2
|
24 |
+
from PIL import Image, ImageDraw, ImageFont
|
25 |
from matplotlib import pyplot as plt
|
26 |
+
|
27 |
import easyocr
|
28 |
from paddleocr import PaddleOCR
|
29 |
+
import supervision as sv
|
30 |
+
import torchvision.transforms as T
|
31 |
+
from torchvision.transforms import ToPILImage
|
32 |
+
from torchvision.ops import box_convert
|
33 |
+
|
34 |
+
# Optional: import AzureOpenAI if used
|
35 |
+
from openai import AzureOpenAI
|
36 |
+
|
37 |
+
# Initialize OCR readers
|
38 |
reader = easyocr.Reader(['en'])
|
39 |
paddle_ocr = PaddleOCR(
|
40 |
+
lang='en', # other languages available
|
41 |
use_angle_cls=False,
|
42 |
+
use_gpu=False, # using cuda might conflict with PyTorch in the same process
|
43 |
show_log=False,
|
44 |
max_batch_size=1024,
|
45 |
use_dilation=True, # improves accuracy
|
46 |
det_db_score_mode='slow', # improves accuracy
|
47 |
+
rec_batch_num=1024
|
48 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
|
50 |
|
51 |
def get_caption_model_processor(model_name, model_name_or_path="Salesforce/blip2-opt-2.7b", device=None):
|
52 |
+
"""
|
53 |
+
Loads the captioning model and processor.
|
54 |
+
Supports either BLIP2 or Florence-2 models.
|
55 |
+
"""
|
56 |
if not device:
|
57 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
58 |
if model_name == "blip2":
|
|
|
60 |
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
|
61 |
if device == 'cpu':
|
62 |
model = Blip2ForConditionalGeneration.from_pretrained(
|
63 |
+
model_name_or_path, device_map=None, torch_dtype=torch.float32
|
64 |
+
)
|
65 |
else:
|
66 |
model = Blip2ForConditionalGeneration.from_pretrained(
|
67 |
+
model_name_or_path, device_map=None, torch_dtype=torch.float16
|
68 |
+
).to(device)
|
69 |
elif model_name == "florence2":
|
70 |
+
from transformers import AutoProcessor, AutoModelForCausalLM
|
71 |
processor = AutoProcessor.from_pretrained("microsoft/Florence-2-base", trust_remote_code=True)
|
72 |
if device == 'cpu':
|
73 |
+
model = AutoModelForCausalLM.from_pretrained(
|
74 |
+
model_name_or_path, torch_dtype=torch.float32, trust_remote_code=True
|
75 |
+
)
|
76 |
else:
|
77 |
+
model = AutoModelForCausalLM.from_pretrained(
|
78 |
+
model_name_or_path, torch_dtype=torch.float16, trust_remote_code=True
|
79 |
+
).to(device)
|
80 |
return {'model': model.to(device), 'processor': processor}
|
81 |
|
82 |
|
83 |
def get_yolo_model(model_path):
|
84 |
+
"""
|
85 |
+
Loads a YOLO model from a given model_path using ultralytics.
|
86 |
+
"""
|
87 |
from ultralytics import YOLO
|
|
|
88 |
model = YOLO(model_path)
|
89 |
return model
|
90 |
|
91 |
|
92 |
@torch.inference_mode()
|
93 |
def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=None, batch_size=32):
|
94 |
+
"""
|
95 |
+
Generates parsed textual content for detected icons from the image.
|
96 |
+
|
97 |
+
Args:
|
98 |
+
filtered_boxes: Tensor of bounding boxes.
|
99 |
+
starting_idx: Starting index for non-OCR boxes.
|
100 |
+
image_source: Original image as a NumPy array.
|
101 |
+
caption_model_processor: Dictionary with keys 'model' and 'processor'.
|
102 |
+
prompt: Optional prompt text.
|
103 |
+
batch_size: Batch size for processing.
|
104 |
+
|
105 |
+
Returns:
|
106 |
+
List of generated texts.
|
107 |
+
"""
|
108 |
to_pil = ToPILImage()
|
109 |
if starting_idx:
|
110 |
non_ocr_boxes = filtered_boxes[starting_idx:]
|
111 |
else:
|
112 |
non_ocr_boxes = filtered_boxes
|
113 |
+
cropped_pil_images = []
|
114 |
+
for coord in non_ocr_boxes:
|
115 |
+
xmin, xmax = int(coord[0] * image_source.shape[1]), int(coord[2] * image_source.shape[1])
|
116 |
+
ymin, ymax = int(coord[1] * image_source.shape[0]), int(coord[3] * image_source.shape[0])
|
117 |
cropped_image = image_source[ymin:ymax, xmin:xmax, :]
|
118 |
+
cropped_pil_images.append(to_pil(cropped_image))
|
119 |
|
120 |
model, processor = caption_model_processor['model'], caption_model_processor['processor']
|
121 |
if not prompt:
|
|
|
126 |
|
127 |
generated_texts = []
|
128 |
device = model.device
|
129 |
+
for i in range(0, len(cropped_pil_images), batch_size):
|
130 |
+
batch = cropped_pil_images[i:i+batch_size]
|
|
|
131 |
if model.device.type == 'cuda':
|
132 |
+
inputs = processor(images=batch, text=[prompt] * len(batch), return_tensors="pt").to(device=device, dtype=torch.float16)
|
133 |
else:
|
134 |
+
inputs = processor(images=batch, text=[prompt] * len(batch), return_tensors="pt").to(device=device)
|
135 |
if 'florence' in model.config.name_or_path:
|
136 |
+
generated_ids = model.generate(
|
137 |
+
input_ids=inputs["input_ids"],
|
138 |
+
pixel_values=inputs["pixel_values"],
|
139 |
+
max_new_tokens=100,
|
140 |
+
num_beams=3,
|
141 |
+
do_sample=False
|
142 |
+
)
|
143 |
else:
|
144 |
+
generated_ids = model.generate(
|
145 |
+
**inputs,
|
146 |
+
max_length=100,
|
147 |
+
num_beams=5,
|
148 |
+
no_repeat_ngram_size=2,
|
149 |
+
early_stopping=True,
|
150 |
+
num_return_sequences=1
|
151 |
+
)
|
152 |
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
|
153 |
generated_text = [gen.strip() for gen in generated_text]
|
154 |
generated_texts.extend(generated_text)
|
|
|
156 |
return generated_texts
|
157 |
|
158 |
|
|
|
159 |
def get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor):
|
160 |
+
"""
|
161 |
+
Generates parsed textual content for detected icons using the phi3_v model variant.
|
162 |
+
"""
|
163 |
to_pil = ToPILImage()
|
164 |
if ocr_bbox:
|
165 |
non_ocr_boxes = filtered_boxes[len(ocr_bbox):]
|
166 |
else:
|
167 |
non_ocr_boxes = filtered_boxes
|
168 |
+
cropped_pil_images = []
|
169 |
+
for coord in non_ocr_boxes:
|
170 |
+
xmin, xmax = int(coord[0] * image_source.shape[1]), int(coord[2] * image_source.shape[1])
|
171 |
+
ymin, ymax = int(coord[1] * image_source.shape[0]), int(coord[3] * image_source.shape[0])
|
172 |
cropped_image = image_source[ymin:ymax, xmin:xmax, :]
|
173 |
+
cropped_pil_images.append(to_pil(cropped_image))
|
174 |
|
175 |
model, processor = caption_model_processor['model'], caption_model_processor['processor']
|
176 |
device = model.device
|
177 |
+
messages = [{"role": "user", "content": "<|image_1|>\ndescribe the icon in one sentence"}]
|
178 |
prompt = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
179 |
|
180 |
batch_size = 5 # Number of samples per batch
|
181 |
generated_texts = []
|
182 |
|
183 |
+
for i in range(0, len(cropped_pil_images), batch_size):
|
184 |
+
images = cropped_pil_images[i:i+batch_size]
|
185 |
image_inputs = [processor.image_processor(x, return_tensors="pt") for x in images]
|
186 |
+
inputs = {'input_ids': [], 'attention_mask': [], 'pixel_values': [], 'image_sizes': []}
|
187 |
texts = [prompt] * len(images)
|
188 |
+
for idx, txt in enumerate(texts):
|
189 |
+
inp = processor._convert_images_texts_to_inputs(image_inputs[idx], txt, return_tensors="pt")
|
190 |
+
inputs['input_ids'].append(inp['input_ids'])
|
191 |
+
inputs['attention_mask'].append(inp['attention_mask'])
|
192 |
+
inputs['pixel_values'].append(inp['pixel_values'])
|
193 |
+
inputs['image_sizes'].append(inp['image_sizes'])
|
194 |
+
max_len = max(x.shape[1] for x in inputs['input_ids'])
|
195 |
+
for idx, v in enumerate(inputs['input_ids']):
|
196 |
+
pad_tensor = processor.tokenizer.pad_token_id * torch.ones(1, max_len - v.shape[1], dtype=torch.long)
|
197 |
+
inputs['input_ids'][idx] = torch.cat([pad_tensor, v], dim=1)
|
198 |
+
pad_att = torch.zeros(1, max_len - v.shape[1], dtype=torch.long)
|
199 |
+
inputs['attention_mask'][idx] = torch.cat([pad_att, inputs['attention_mask'][idx]], dim=1)
|
200 |
inputs_cat = {k: torch.concatenate(v).to(device) for k, v in inputs.items()}
|
201 |
|
202 |
+
generation_args = {
|
203 |
+
"max_new_tokens": 25,
|
204 |
+
"temperature": 0.01,
|
205 |
+
"do_sample": False,
|
206 |
+
}
|
207 |
+
generate_ids = model.generate(**inputs_cat, eos_token_id=processor.tokenizer.eos_token_id, **generation_args)
|
208 |
+
# Remove input tokens from the generated sequence
|
209 |
generate_ids = generate_ids[:, inputs_cat['input_ids'].shape[1]:]
|
210 |
response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
211 |
response = [res.strip('\n').strip() for res in response]
|
|
|
213 |
|
214 |
return generated_texts
|
215 |
|
216 |
+
|
217 |
def remove_overlap(boxes, iou_threshold, ocr_bbox=None):
|
218 |
+
"""
|
219 |
+
Removes overlapping bounding boxes based on IoU and optionally considers OCR boxes.
|
220 |
+
|
221 |
+
Args:
|
222 |
+
boxes: Tensor of bounding boxes (in xyxy format).
|
223 |
+
iou_threshold: IoU threshold to determine overlaps.
|
224 |
+
ocr_bbox: Optional list of OCR bounding boxes.
|
225 |
+
|
226 |
+
Returns:
|
227 |
+
Filtered boxes as a torch.Tensor.
|
228 |
+
"""
|
229 |
assert ocr_bbox is None or isinstance(ocr_bbox, List)
|
230 |
|
231 |
def box_area(box):
|
|
|
239 |
return max(0, x2 - x1) * max(0, y2 - y1)
|
240 |
|
241 |
def IoU(box1, box2):
|
242 |
+
inter = intersection_area(box1, box2)
|
243 |
+
union = box_area(box1) + box_area(box2) - inter + 1e-6
|
244 |
+
ratio1 = inter / box_area(box1) if box_area(box1) > 0 else 0
|
245 |
+
ratio2 = inter / box_area(box2) if box_area(box2) > 0 else 0
|
246 |
+
return max(inter / union, ratio1, ratio2)
|
|
|
|
|
|
|
247 |
|
248 |
def is_inside(box1, box2):
|
249 |
+
inter = intersection_area(box1, box2)
|
250 |
+
return (inter / box_area(box1)) > 0.95
|
|
|
|
|
251 |
|
252 |
boxes = boxes.tolist()
|
253 |
filtered_boxes = []
|
254 |
if ocr_bbox:
|
255 |
filtered_boxes.extend(ocr_bbox)
|
|
|
256 |
for i, box1 in enumerate(boxes):
|
|
|
257 |
is_valid_box = True
|
258 |
for j, box2 in enumerate(boxes):
|
|
|
259 |
if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
|
260 |
is_valid_box = False
|
261 |
break
|
262 |
if is_valid_box:
|
|
|
263 |
if ocr_bbox:
|
264 |
+
# Only add the box if it does not overlap with any OCR box
|
265 |
+
if not any(IoU(box1, box3) > iou_threshold and not is_inside(box1, box3) for box3 in ocr_bbox):
|
266 |
filtered_boxes.append(box1)
|
267 |
else:
|
268 |
filtered_boxes.append(box1)
|
|
|
270 |
|
271 |
|
272 |
def remove_overlap_new(boxes, iou_threshold, ocr_bbox=None):
|
273 |
+
"""
|
274 |
+
Removes overlapping boxes with OCR priority.
|
275 |
+
|
276 |
+
Args:
|
277 |
+
boxes: List of dictionaries, each with keys: 'type', 'bbox', 'interactivity', 'content'.
|
278 |
+
iou_threshold: IoU threshold for removal.
|
279 |
+
ocr_bbox: List of OCR box dictionaries.
|
280 |
+
|
281 |
+
Returns:
|
282 |
+
A list of filtered box dictionaries.
|
283 |
+
"""
|
284 |
assert ocr_bbox is None or isinstance(ocr_bbox, List)
|
285 |
|
286 |
def box_area(box):
|
|
|
294 |
return max(0, x2 - x1) * max(0, y2 - y1)
|
295 |
|
296 |
def IoU(box1, box2):
|
297 |
+
inter = intersection_area(box1, box2)
|
298 |
+
union = box_area(box1) + box_area(box2) - inter + 1e-6
|
299 |
+
ratio1 = inter / box_area(box1) if box_area(box1) > 0 else 0
|
300 |
+
ratio2 = inter / box_area(box2) if box_area(box2) > 0 else 0
|
301 |
+
return max(inter / union, ratio1, ratio2)
|
|
|
|
|
|
|
302 |
|
303 |
def is_inside(box1, box2):
|
304 |
+
inter = intersection_area(box1, box2)
|
305 |
+
return (inter / box_area(box1)) > 0.80
|
|
|
|
|
306 |
|
|
|
307 |
filtered_boxes = []
|
308 |
if ocr_bbox:
|
309 |
filtered_boxes.extend(ocr_bbox)
|
|
|
310 |
for i, box1_elem in enumerate(boxes):
|
311 |
box1 = box1_elem['bbox']
|
312 |
is_valid_box = True
|
313 |
for j, box2_elem in enumerate(boxes):
|
|
|
314 |
box2 = box2_elem['bbox']
|
315 |
if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
|
316 |
is_valid_box = False
|
317 |
break
|
318 |
if is_valid_box:
|
|
|
319 |
if ocr_bbox:
|
|
|
320 |
box_added = False
|
321 |
for box3_elem in ocr_bbox:
|
322 |
+
box3 = box3_elem['bbox']
|
323 |
+
if is_inside(box3, box1):
|
324 |
+
try:
|
325 |
+
filtered_boxes.append({
|
326 |
+
'type': 'text',
|
327 |
+
'bbox': box1_elem['bbox'],
|
328 |
+
'interactivity': True,
|
329 |
+
'content': box3_elem['content']
|
330 |
+
})
|
331 |
+
filtered_boxes.remove(box3_elem)
|
332 |
+
except Exception:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
333 |
continue
|
334 |
+
elif is_inside(box1, box3):
|
335 |
+
box_added = True
|
336 |
+
break
|
337 |
if not box_added:
|
338 |
+
filtered_boxes.append({
|
339 |
+
'type': 'icon',
|
340 |
+
'bbox': box1_elem['bbox'],
|
341 |
+
'interactivity': True,
|
342 |
+
'content': None
|
343 |
+
})
|
344 |
else:
|
345 |
filtered_boxes.append(box1)
|
346 |
+
return filtered_boxes # Optionally, you could return torch.tensor(filtered_boxes) if needed
|
347 |
|
348 |
|
349 |
def load_image(image_path: str) -> Tuple[np.array, torch.Tensor]:
|
350 |
+
"""
|
351 |
+
Loads an image and applies transformations.
|
352 |
+
|
353 |
+
Returns:
|
354 |
+
image: Original image as a NumPy array.
|
355 |
+
image_transformed: Transformed tensor.
|
356 |
+
"""
|
357 |
+
transform = T.Compose([
|
358 |
+
T.RandomResize([800], max_size=1333),
|
359 |
+
T.ToTensor(),
|
360 |
+
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
|
361 |
+
])
|
362 |
image_source = Image.open(image_path).convert("RGB")
|
363 |
image = np.asarray(image_source)
|
364 |
image_transformed, _ = transform(image_source, None)
|
365 |
return image, image_transformed
|
366 |
|
367 |
|
368 |
+
def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor, phrases: List[str],
|
369 |
+
text_scale: float, text_padding=5, text_thickness=2, thickness=3) -> Tuple[np.ndarray, dict]:
|
370 |
+
"""
|
371 |
+
Annotates an image with bounding boxes and labels.
|
372 |
+
|
373 |
+
Args:
|
374 |
+
image_source: Source image as a NumPy array.
|
375 |
+
boxes: Bounding boxes in cxcywh format (normalized).
|
376 |
+
logits: Confidence scores for each bounding box.
|
377 |
+
phrases: List of labels.
|
378 |
+
text_scale, text_padding, text_thickness, thickness: Annotation parameters.
|
379 |
+
|
380 |
Returns:
|
381 |
+
Annotated image and a dictionary of label coordinates.
|
382 |
"""
|
383 |
h, w, _ = image_source.shape
|
384 |
boxes = boxes * torch.Tensor([w, h, w, h])
|
|
|
388 |
|
389 |
labels = [f"{phrase}" for phrase in range(boxes.shape[0])]
|
390 |
|
391 |
+
# Import the custom box annotator from your project structure.
|
392 |
+
from util.box_annotator import BoxAnnotator
|
393 |
+
box_annotator = BoxAnnotator(text_scale=text_scale, text_padding=text_padding,
|
394 |
+
text_thickness=text_thickness, thickness=thickness)
|
395 |
annotated_frame = image_source.copy()
|
396 |
+
annotated_frame = box_annotator.annotate(scene=annotated_frame, detections=detections, labels=labels, image_size=(w, h))
|
397 |
|
398 |
label_coordinates = {f"{phrase}": v for phrase, v in zip(phrases, xywh)}
|
399 |
return annotated_frame, label_coordinates
|
400 |
|
401 |
|
402 |
def predict(model, image, caption, box_threshold, text_threshold):
|
|
|
403 |
"""
|
404 |
+
Uses a Hugging Face model to perform grounded object detection.
|
405 |
+
|
406 |
+
Args:
|
407 |
+
model: Dictionary with 'model' and 'processor'.
|
408 |
+
image: Input PIL image.
|
409 |
+
caption: Caption text.
|
410 |
+
box_threshold: Confidence threshold for boxes.
|
411 |
+
text_threshold: Threshold for text detection.
|
412 |
+
|
413 |
+
Returns:
|
414 |
+
boxes, logits, phrases from the detection.
|
415 |
+
"""
|
416 |
+
model_obj, processor = model['model'], model['processor']
|
417 |
+
device = model_obj.device
|
418 |
|
419 |
inputs = processor(images=image, text=caption, return_tensors="pt").to(device)
|
420 |
with torch.no_grad():
|
421 |
+
outputs = model_obj(**inputs)
|
422 |
|
423 |
results = processor.post_process_grounded_object_detection(
|
424 |
outputs,
|
425 |
inputs.input_ids,
|
426 |
+
box_threshold=box_threshold,
|
427 |
+
text_threshold=text_threshold,
|
428 |
target_sizes=[image.size[::-1]]
|
429 |
)[0]
|
430 |
boxes, logits, phrases = results["boxes"], results["scores"], results["labels"]
|
|
|
432 |
|
433 |
|
434 |
def predict_yolo(model, image_path, box_threshold, imgsz, scale_img, iou_threshold=0.7):
|
435 |
+
"""
|
436 |
+
Uses a YOLO model for object detection.
|
437 |
+
|
438 |
+
Args:
|
439 |
+
model: YOLO model instance.
|
440 |
+
image_path: Path to the image.
|
441 |
+
box_threshold: Confidence threshold.
|
442 |
+
imgsz: Image size for scaling (if scale_img is True).
|
443 |
+
scale_img: Boolean flag to scale the image.
|
444 |
+
iou_threshold: IoU threshold for non-max suppression.
|
445 |
+
|
446 |
+
Returns:
|
447 |
+
Bounding boxes, confidence scores, and placeholder phrases.
|
448 |
+
"""
|
449 |
kwargs = {
|
450 |
+
'conf': box_threshold, # Confidence threshold
|
451 |
+
'iou': iou_threshold, # IoU threshold
|
452 |
'verbose': False
|
453 |
}
|
|
|
454 |
if scale_img:
|
455 |
kwargs['imgsz'] = imgsz
|
456 |
+
|
457 |
results = model.predict(image_path, **kwargs)
|
458 |
boxes = results[0].boxes.xyxy
|
459 |
conf = results[0].boxes.conf
|
460 |
return boxes, conf, [str(i) for i in range(len(boxes))]
|
461 |
|
462 |
|
463 |
+
def get_som_labeled_img(img_path, model=None, BOX_TRESHOLD=0.01, output_coord_in_ratio=False, ocr_bbox=None,
|
464 |
+
text_scale=0.4, text_padding=5, draw_bbox_config=None, caption_model_processor=None,
|
465 |
+
ocr_text=[], use_local_semantics=True, iou_threshold=0.9, prompt=None, scale_img=False,
|
466 |
+
imgsz=None, batch_size=None):
|
467 |
+
"""
|
468 |
+
Processes an image to generate semantic (SOM) labels.
|
469 |
+
|
470 |
+
Args:
|
471 |
+
img_path: Path to the image.
|
472 |
+
model: YOLO model for detection.
|
473 |
+
BOX_TRESHOLD: Confidence threshold for box prediction.
|
474 |
+
output_coord_in_ratio: If True, output coordinates in ratio.
|
475 |
+
ocr_bbox: OCR bounding boxes.
|
476 |
+
text_scale, text_padding: Parameters for drawing annotations.
|
477 |
+
draw_bbox_config: Custom configuration for bounding box drawing.
|
478 |
+
caption_model_processor: Dictionary with caption model and processor.
|
479 |
+
ocr_text: List of OCR-detected texts.
|
480 |
+
use_local_semantics: Whether to use local semantic processing.
|
481 |
+
iou_threshold: IoU threshold for filtering overlaps.
|
482 |
+
prompt: Optional caption prompt.
|
483 |
+
scale_img: Whether to scale the image.
|
484 |
+
imgsz: Image size for YOLO.
|
485 |
+
batch_size: Batch size for captioning.
|
486 |
+
|
487 |
+
Returns:
|
488 |
+
Encoded annotated image, label coordinates, and filtered boxes.
|
489 |
"""
|
490 |
image_source = Image.open(img_path).convert("RGB")
|
491 |
w, h = image_source.size
|
492 |
if not imgsz:
|
493 |
imgsz = (h, w)
|
494 |
+
# Run YOLO detection
|
495 |
+
xyxy, logits, phrases = predict_yolo(
|
496 |
+
model=model, image_path=img_path, box_threshold=BOX_TRESHOLD,
|
497 |
+
imgsz=imgsz, scale_img=scale_img, iou_threshold=0.1
|
498 |
+
)
|
499 |
xyxy = xyxy / torch.Tensor([w, h, w, h]).to(xyxy.device)
|
500 |
+
image_source_np = np.asarray(image_source)
|
501 |
phrases = [str(i) for i in range(len(phrases))]
|
502 |
|
503 |
+
# Process OCR bounding boxes (if any)
|
|
|
504 |
if ocr_bbox:
|
505 |
ocr_bbox = torch.tensor(ocr_bbox) / torch.Tensor([w, h, w, h])
|
506 |
+
ocr_bbox = ocr_bbox.tolist()
|
507 |
else:
|
508 |
print('no ocr bbox!!!')
|
509 |
ocr_bbox = None
|
|
|
|
|
|
|
510 |
|
511 |
+
ocr_bbox_elem = [{'type': 'text', 'bbox': box, 'interactivity': False, 'content': txt}
|
512 |
+
for box, txt in zip(ocr_bbox, ocr_text)]
|
513 |
+
xyxy_elem = [{'type': 'icon', 'bbox': box, 'interactivity': True, 'content': None}
|
514 |
+
for box in xyxy.tolist()]
|
515 |
filtered_boxes = remove_overlap_new(boxes=xyxy_elem, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox_elem)
|
516 |
|
517 |
+
# Sort filtered boxes so that boxes with 'content' == None are at the end
|
518 |
filtered_boxes_elem = sorted(filtered_boxes, key=lambda x: x['content'] is None)
|
|
|
519 |
starting_idx = next((i for i, box in enumerate(filtered_boxes_elem) if box['content'] is None), -1)
|
520 |
+
filtered_boxes_tensor = torch.tensor([box['bbox'] for box in filtered_boxes_elem])
|
521 |
|
522 |
+
# Generate parsed icon semantics if required
|
|
|
523 |
if use_local_semantics:
|
524 |
caption_model = caption_model_processor['model']
|
525 |
+
if 'phi3_v' in caption_model.config.model_type:
|
526 |
+
parsed_content_icon = get_parsed_content_icon_phi3v(filtered_boxes_tensor, ocr_bbox, image_source_np, caption_model_processor)
|
527 |
else:
|
528 |
+
parsed_content_icon = get_parsed_content_icon(filtered_boxes_tensor, starting_idx, image_source_np, caption_model_processor, prompt=prompt, batch_size=batch_size)
|
529 |
ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
|
530 |
icon_start = len(ocr_text)
|
531 |
parsed_content_icon_ls = []
|
532 |
+
# Fill boxes with no OCR content with parsed icon content
|
533 |
+
for box in filtered_boxes_elem:
|
534 |
+
if box['content'] is None and parsed_content_icon:
|
535 |
box['content'] = parsed_content_icon.pop(0)
|
536 |
for i, txt in enumerate(parsed_content_icon):
|
537 |
parsed_content_icon_ls.append(f"Icon Box ID {str(i+icon_start)}: {txt}")
|
|
|
540 |
ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
|
541 |
parsed_content_merged = ocr_text
|
542 |
|
543 |
+
filtered_boxes_cxcywh = box_convert(boxes=filtered_boxes_tensor, in_fmt="xyxy", out_fmt="cxcywh")
|
544 |
+
phrases = [i for i in range(len(filtered_boxes_cxcywh))]
|
|
|
545 |
|
546 |
+
# Annotate image with bounding boxes and labels
|
547 |
if draw_bbox_config:
|
548 |
+
annotated_frame, label_coordinates = annotate(
|
549 |
+
image_source=image_source_np, boxes=filtered_boxes_cxcywh, logits=logits, phrases=phrases, **draw_bbox_config
|
550 |
+
)
|
551 |
else:
|
552 |
+
annotated_frame, label_coordinates = annotate(
|
553 |
+
image_source=image_source_np, boxes=filtered_boxes_cxcywh, logits=logits, phrases=phrases,
|
554 |
+
text_scale=text_scale, text_padding=text_padding
|
555 |
+
)
|
556 |
|
557 |
pil_img = Image.fromarray(annotated_frame)
|
558 |
buffered = io.BytesIO()
|
559 |
pil_img.save(buffered, format="PNG")
|
560 |
encoded_image = base64.b64encode(buffered.getvalue()).decode('ascii')
|
561 |
+
|
562 |
if output_coord_in_ratio:
|
563 |
+
label_coordinates = {k: [v[0] / w, v[1] / h, v[2] / w, v[3] / h] for k, v in label_coordinates.items()}
|
|
|
564 |
assert w == annotated_frame.shape[1] and h == annotated_frame.shape[0]
|
565 |
|
566 |
return encoded_image, label_coordinates, filtered_boxes_elem
|
567 |
|
568 |
|
569 |
def get_xywh(input):
|
570 |
+
"""
|
571 |
+
Converts a bounding box from a list of two points into (x, y, width, height).
|
572 |
+
"""
|
573 |
+
x, y = input[0][0], input[0][1]
|
574 |
+
w = input[2][0] - input[0][0]
|
575 |
+
h = input[2][1] - input[0][1]
|
576 |
+
return int(x), int(y), int(w), int(h)
|
577 |
+
|
578 |
|
579 |
def get_xyxy(input):
|
580 |
+
"""
|
581 |
+
Converts a bounding box from a list of two points into (x, y, x2, y2).
|
582 |
+
"""
|
583 |
+
x, y = input[0][0], input[0][1]
|
584 |
+
x2, y2 = input[2][0], input[2][1]
|
585 |
+
return int(x), int(y), int(x2), int(y2)
|
586 |
+
|
587 |
|
588 |
def get_xywh_yolo(input):
|
589 |
+
"""
|
590 |
+
Converts a YOLO-style bounding box (x1, y1, x2, y2) into (x, y, width, height).
|
591 |
+
"""
|
592 |
+
x, y = input[0], input[1]
|
593 |
+
w = input[2] - input[0]
|
594 |
+
h = input[3] - input[1]
|
595 |
+
return int(x), int(y), int(w), int(h)
|
596 |
|
597 |
|
598 |
+
def check_ocr_box(image_path, display_img=True, output_bb_format='xywh', goal_filtering=None, easyocr_args=None, use_paddleocr=False):
|
599 |
+
"""
|
600 |
+
Runs OCR on the given image using PaddleOCR or EasyOCR and optionally displays annotated results.
|
601 |
+
|
602 |
+
Returns:
|
603 |
+
A tuple containing:
|
604 |
+
- A tuple (text, bounding boxes)
|
605 |
+
- The goal_filtering parameter (unchanged)
|
606 |
+
"""
|
607 |
if use_paddleocr:
|
608 |
+
text_threshold = 0.5 if easyocr_args is None else easyocr_args.get('text_threshold', 0.5)
|
|
|
|
|
|
|
609 |
result = paddle_ocr.ocr(image_path, cls=False)[0]
|
610 |
conf = [item[1] for item in result]
|
611 |
coord = [item[0] for item in result if item[1][1] > text_threshold]
|
|
|
614 |
if easyocr_args is None:
|
615 |
easyocr_args = {}
|
616 |
result = reader.readtext(image_path, **easyocr_args)
|
|
|
617 |
coord = [item[0] for item in result]
|
618 |
text = [item[1] for item in result]
|
619 |
+
|
620 |
if display_img:
|
621 |
opencv_img = cv2.imread(image_path)
|
622 |
opencv_img = cv2.cvtColor(opencv_img, cv2.COLOR_RGB2BGR)
|
623 |
bb = []
|
624 |
for item in coord:
|
625 |
x, y, a, b = get_xywh(item)
|
|
|
626 |
bb.append((x, y, a, b))
|
627 |
+
cv2.rectangle(opencv_img, (x, y), (x + a, y + b), (0, 255, 0), 2)
|
|
|
|
|
628 |
plt.imshow(opencv_img)
|
629 |
else:
|
630 |
if output_bb_format == 'xywh':
|
631 |
bb = [get_xywh(item) for item in coord]
|
632 |
elif output_bb_format == 'xyxy':
|
633 |
bb = [get_xyxy(item) for item in coord]
|
|
|
634 |
return (text, bb), goal_filtering
|
|
|
|