tappyness1
commited on
Commit
·
264e65b
1
Parent(s):
a1027a2
default to side-by-side
Browse files- src/pred_analysis_STEE.py +0 -595
- src/st_image_tools.py +2 -2
src/pred_analysis_STEE.py
DELETED
@@ -1,595 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
import re
|
3 |
-
import cv2
|
4 |
-
import time
|
5 |
-
|
6 |
-
import numpy as np
|
7 |
-
import pandas as pd
|
8 |
-
import xml.etree.ElementTree as ET
|
9 |
-
|
10 |
-
from pathlib import Path
|
11 |
-
from torchvision import transforms
|
12 |
-
from configparser import ConfigParser, ExtendedInterpolation
|
13 |
-
from ast import literal_eval
|
14 |
-
|
15 |
-
from src.models.model import Model
|
16 |
-
from src.models.eval.confusion_matrix import ConfusionMatrix
|
17 |
-
|
18 |
-
|
19 |
-
def generate_inference_from_img_folder(csv_file, model_cfg, img_folder, ckpt_file,
|
20 |
-
nms_thresh, conf_thresh, device="cuda" ,csv_path=None):
|
21 |
-
"""[Retrieve the inference information of the test images given a model checkpoint trained]
|
22 |
-
|
23 |
-
Parameters
|
24 |
-
----------
|
25 |
-
csv_file : [str]
|
26 |
-
[path of the csv file containing the information of the test images]
|
27 |
-
model_cfg : [str]
|
28 |
-
[path of the model config file to use, specific to the checkpoint file]
|
29 |
-
img_folder : [str]
|
30 |
-
[folder containing the images]
|
31 |
-
ckpt_file : [str]
|
32 |
-
[path of the model checkpoint file to use for model inference]
|
33 |
-
nms_thresh : [float]
|
34 |
-
[Non-maximum suppression threshold to use for the model inference, values between 0 to 1]
|
35 |
-
conf_thresh : [float]
|
36 |
-
[Confidence threshold to use for the model inference, values between 0 to 1]
|
37 |
-
device : str, optional
|
38 |
-
[device to use for inference, option: "cuda" or "cpu"], by default "cuda"
|
39 |
-
csv_path : [str], optional
|
40 |
-
[path to save the pandas.DataFrame output as a csv], by default None i.e. csv not generated
|
41 |
-
|
42 |
-
Returns
|
43 |
-
-------
|
44 |
-
df : [pandas.DataFrame]
|
45 |
-
[dataframe containing the inference information of the test images]
|
46 |
-
"""
|
47 |
-
|
48 |
-
pl_config = ConfigParser(interpolation=ExtendedInterpolation())
|
49 |
-
pl_config.read(model_cfg)
|
50 |
-
|
51 |
-
model_selected = Model(pl_config)
|
52 |
-
|
53 |
-
df_original = pd.read_csv(csv_file)
|
54 |
-
# Only perform inference on test images with at least 1 ground truth.
|
55 |
-
df_test = df_original[df_original['remarks_xml'] == 'Available xml file'].reset_index()
|
56 |
-
df_test = df_test[df_test['set_type'] == 'Test'].reset_index()
|
57 |
-
|
58 |
-
img_number = 0
|
59 |
-
prediction_info_list = []
|
60 |
-
for _,rows in df_test.iterrows():
|
61 |
-
img_file = rows["image_file_name"]
|
62 |
-
img_number += 1
|
63 |
-
inference_start_time = time.time()
|
64 |
-
img_file_path = os.path.join(img_folder,img_file)
|
65 |
-
|
66 |
-
# Perform inference on image with ckpt file with device either "cuda" or "cpu"
|
67 |
-
# img_inference = model_selected.inference(device='cpu', img_path=img_file_path, ckpt_path=ckpt_file)
|
68 |
-
img_inference = model_selected.inference(
|
69 |
-
device=device, img_path=img_file_path, ckpt_path=ckpt_file, nms_thresh=nms_thresh, conf_thresh=conf_thresh)
|
70 |
-
|
71 |
-
# Sieve out inference
|
72 |
-
predicted_boxes_unsorted = img_inference[0].tolist()
|
73 |
-
predicted_labels_unsorted = img_inference[1].tolist()
|
74 |
-
predicted_confidence_unsorted = img_inference[2].tolist()
|
75 |
-
|
76 |
-
# print(f"Pre Boxes: {predicted_boxes}")
|
77 |
-
# print(f"Pre Labels: {predicted_labels}")
|
78 |
-
# print(f"Pre Labels: {predicted_confidence}")
|
79 |
-
|
80 |
-
# Sorting input
|
81 |
-
predicted_boxes = [x for _,x in sorted(zip(predicted_confidence_unsorted,predicted_boxes_unsorted), reverse=True)]
|
82 |
-
predicted_labels = [x for _,x in sorted(zip(predicted_confidence_unsorted,predicted_labels_unsorted), reverse=True)]
|
83 |
-
predicted_confidence = sorted(predicted_confidence_unsorted, reverse=True)
|
84 |
-
|
85 |
-
# print(f"Post Boxes: {predicted_boxes}")
|
86 |
-
# print(f"Post Labels: {predicted_labels}")
|
87 |
-
# print(f"Post Labels: {predicted_confidence}")
|
88 |
-
|
89 |
-
predicted_boxes_int = []
|
90 |
-
for box in predicted_boxes:
|
91 |
-
box_int = [round(x) for x in box]
|
92 |
-
predicted_boxes_int.append(box_int)
|
93 |
-
|
94 |
-
# Prepare inputs for confusion matrix
|
95 |
-
cm_detections_list = []
|
96 |
-
for prediction in range(len(predicted_boxes)):
|
97 |
-
detection_list = predicted_boxes[prediction]
|
98 |
-
detection_list.append(predicted_confidence[prediction])
|
99 |
-
detection_list.append(predicted_labels[prediction])
|
100 |
-
cm_detections_list.append(detection_list)
|
101 |
-
|
102 |
-
# Re generate predicted boxes
|
103 |
-
predicted_boxes = [x for _,x in sorted(zip(predicted_confidence_unsorted,predicted_boxes_unsorted), reverse=True)]
|
104 |
-
|
105 |
-
inference_time_per_image = round(time.time() - inference_start_time, 2)
|
106 |
-
if img_number%100 == 0:
|
107 |
-
print(f'Performing inference on Image {img_number}: {img_file_path}')
|
108 |
-
print(f'Time taken for image: {inference_time_per_image}')
|
109 |
-
|
110 |
-
prediction_info = {
|
111 |
-
"image_file_path": img_file_path,
|
112 |
-
"image_file_name": img_file,
|
113 |
-
"number_of_predictions": len(predicted_boxes),
|
114 |
-
"predicted_boxes": predicted_boxes,
|
115 |
-
"predicted_boxes_int": predicted_boxes_int,
|
116 |
-
"predicted_labels": predicted_labels,
|
117 |
-
"predicted_confidence": predicted_confidence,
|
118 |
-
"cm_detections_list": cm_detections_list,
|
119 |
-
"inference_time": inference_time_per_image
|
120 |
-
}
|
121 |
-
prediction_info_list.append(prediction_info)
|
122 |
-
|
123 |
-
df = pd.DataFrame(prediction_info_list)
|
124 |
-
|
125 |
-
if csv_path is not None:
|
126 |
-
df.to_csv(csv_path, index=False)
|
127 |
-
print ("Dataframe saved as csv to " + csv_path)
|
128 |
-
|
129 |
-
return df
|
130 |
-
|
131 |
-
def get_gt_from_img_folder(csv_file, img_folder, xml_folder, names_file, map_start_index=1, csv_path=None):
|
132 |
-
"""[Retrieve the ground truth information of the test images]
|
133 |
-
|
134 |
-
Parameters
|
135 |
-
----------
|
136 |
-
csv_file : [str]
|
137 |
-
[path of the csv file containing the information of the test images]
|
138 |
-
img_folder : [str]
|
139 |
-
[folder containing the images]
|
140 |
-
xml_folder : [str]
|
141 |
-
[folder containing the xml files associated with the images]
|
142 |
-
names_file : [str]
|
143 |
-
[names file containing the class labels of interest]
|
144 |
-
map_start_index : int, optional
|
145 |
-
[attach a number to each class label listed in names file, starting from number given by map_start_index], by default 1
|
146 |
-
csv_path : [str], optional
|
147 |
-
[path to save the pandas.DataFrame output as a csv], by default None i.e. csv not generated
|
148 |
-
|
149 |
-
Returns
|
150 |
-
-------
|
151 |
-
df : [pandas.DataFrame]
|
152 |
-
[dataframe containing the ground truth information of the test images]
|
153 |
-
"""
|
154 |
-
|
155 |
-
df_original = pd.read_csv(csv_file)
|
156 |
-
|
157 |
-
# Only perform inference on test images with at least 1 ground truth.
|
158 |
-
df_test = df_original[df_original['remarks_xml'] == 'Available xml file'].reset_index()
|
159 |
-
df_test = df_test[df_test['set_type'] == 'Test'].reset_index()
|
160 |
-
|
161 |
-
# Create a dictionary to map numeric class as class labels
|
162 |
-
class_labels_dict = {}
|
163 |
-
with open(names_file) as f:
|
164 |
-
for index,line in enumerate(f):
|
165 |
-
idx = index + map_start_index
|
166 |
-
class_labels = line.splitlines()[0]
|
167 |
-
class_labels_dict[class_labels] = idx
|
168 |
-
|
169 |
-
gt_info_list = []
|
170 |
-
# for img_file in os.listdir(img_folder):
|
171 |
-
# if re.search(".jpg", img_file):
|
172 |
-
for _,rows in df_test.iterrows():
|
173 |
-
img_file = rows["image_file_name"]
|
174 |
-
# file_stem = Path(img_file_path).stem
|
175 |
-
|
176 |
-
# Get img tensor
|
177 |
-
img_file_path = os.path.join(img_folder,img_file)
|
178 |
-
img = cv2.imread(filename = img_file_path)
|
179 |
-
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
180 |
-
|
181 |
-
# Get associated xml file
|
182 |
-
file_stem = Path(img_file_path).stem
|
183 |
-
xml_file_path = xml_folder + file_stem + ".xml"
|
184 |
-
|
185 |
-
tree = ET.parse(xml_file_path)
|
186 |
-
root = tree.getroot()
|
187 |
-
|
188 |
-
for image_detail in root.findall('size'):
|
189 |
-
image_width = float(image_detail.find('width').text)
|
190 |
-
image_height = float(image_detail.find('height').text)
|
191 |
-
|
192 |
-
class_index_list = []
|
193 |
-
bb_list = []
|
194 |
-
truncated_list = []
|
195 |
-
occluded_list = []
|
196 |
-
for item in root.findall('object'):
|
197 |
-
if item.find('truncated') is not None:
|
198 |
-
truncated = int(item.find('truncated').text)
|
199 |
-
else:
|
200 |
-
truncated = 0
|
201 |
-
|
202 |
-
if item.find('occluded').text is not None:
|
203 |
-
occluded = int(item.find('occluded').text)
|
204 |
-
else:
|
205 |
-
occluded = 0
|
206 |
-
|
207 |
-
for bb_details in item.findall('bndbox'):
|
208 |
-
class_label = item.find('name').text
|
209 |
-
class_index = class_labels_dict[class_label]
|
210 |
-
xmin = float(bb_details.find('xmin').text)
|
211 |
-
ymin = float(bb_details.find('ymin').text)
|
212 |
-
xmax = float(bb_details.find('xmax').text)
|
213 |
-
ymax = float(bb_details.find('ymax').text)
|
214 |
-
|
215 |
-
class_index_list.append(class_index)
|
216 |
-
bb_list.append([xmin,ymin,xmax,ymax])
|
217 |
-
truncated_list.append(truncated)
|
218 |
-
occluded_list.append(occluded)
|
219 |
-
|
220 |
-
transform = A.Compose([
|
221 |
-
A.Resize(608,608),
|
222 |
-
ToTensor()
|
223 |
-
],
|
224 |
-
bbox_params=A.BboxParams(format='pascal_voc',
|
225 |
-
label_fields=['class_labels']),
|
226 |
-
)
|
227 |
-
|
228 |
-
augmented = transform(image=img, bboxes = bb_list, class_labels = class_index_list)
|
229 |
-
# img comes out as int, need to change to float.
|
230 |
-
img = augmented['image'].float()
|
231 |
-
gt_boxes = augmented['bboxes']
|
232 |
-
gt_boxes_list = [list(box) for box in gt_boxes]
|
233 |
-
gt_labels = augmented['class_labels']
|
234 |
-
|
235 |
-
gt_boxes_int = []
|
236 |
-
for box in gt_boxes:
|
237 |
-
box_int = [round(x) for x in box]
|
238 |
-
gt_boxes_int.append(box_int)
|
239 |
-
|
240 |
-
cm_gt_list = []
|
241 |
-
for gt in range(len(gt_boxes)):
|
242 |
-
gt_list = [gt_labels[gt]]
|
243 |
-
gt_list.extend(gt_boxes[gt])
|
244 |
-
cm_gt_list.append(gt_list)
|
245 |
-
|
246 |
-
# Calculate and Group by Size of Ground Truth
|
247 |
-
gt_area_list = []
|
248 |
-
gt_area_type = []
|
249 |
-
for gt_box in gt_boxes:
|
250 |
-
gt_area = (gt_box[3] - gt_box[1]) * (gt_box[2] - gt_box[0])
|
251 |
-
gt_area_list.append(gt_area)
|
252 |
-
|
253 |
-
if gt_area < 32*32:
|
254 |
-
area_type = "S"
|
255 |
-
gt_area_type.append(area_type)
|
256 |
-
elif gt_area < 96*96:
|
257 |
-
area_type = "M"
|
258 |
-
gt_area_type.append(area_type)
|
259 |
-
else:
|
260 |
-
area_type = "L"
|
261 |
-
gt_area_type.append(area_type)
|
262 |
-
|
263 |
-
gt_info = {
|
264 |
-
"image_file_path": img_file_path,
|
265 |
-
"image_file_name": img_file,
|
266 |
-
"image_width": image_width,
|
267 |
-
"image_height": image_height,
|
268 |
-
"number_of_gt": len(gt_boxes_list),
|
269 |
-
"gt_labels": gt_labels,
|
270 |
-
"gt_boxes": gt_boxes_list,
|
271 |
-
"gt_boxes_int": gt_boxes_int,
|
272 |
-
"cm_gt_list": cm_gt_list,
|
273 |
-
"gt_area_list": gt_area_list,
|
274 |
-
"gt_area_type": gt_area_type,
|
275 |
-
"truncated_list": truncated_list,
|
276 |
-
"occluded_list": occluded_list
|
277 |
-
}
|
278 |
-
gt_info_list.append(gt_info)
|
279 |
-
|
280 |
-
df = pd.DataFrame(gt_info_list)
|
281 |
-
|
282 |
-
if csv_path is not None:
|
283 |
-
df.to_csv(csv_path, index=False)
|
284 |
-
print ("Dataframe saved as csv to " + csv_path)
|
285 |
-
|
286 |
-
return df
|
287 |
-
|
288 |
-
def combine_gt_predictions(csv_file, img_folder, xml_folder, names_file, model_cfg, ckpt_file, csv_save_folder,
|
289 |
-
device="cuda", nms_threshold=0.1, confidence_threshold=0.7, iou_threshold=0.4, gt_statistics=True):
|
290 |
-
"""[Retrieve the combined inference and ground truth information of the test images]
|
291 |
-
|
292 |
-
Parameters
|
293 |
-
----------
|
294 |
-
csv_file : [str]
|
295 |
-
[path of the csv file containing the information of the test images]
|
296 |
-
img_folder : [str]
|
297 |
-
[folder containing the images]
|
298 |
-
xml_folder : [str]
|
299 |
-
[folder containing the xml files associated with the images]
|
300 |
-
names_file : [str]
|
301 |
-
[names file containing the class labels of interest]
|
302 |
-
model_cfg : [str]
|
303 |
-
[path of the model config file to use, specific to the checkpoint file]
|
304 |
-
ckpt_file : [str]
|
305 |
-
[path of the model checkpoint file to use for model inference]
|
306 |
-
csv_save_folder : [str]
|
307 |
-
[folder to save the generated csv files]
|
308 |
-
device : str, optional
|
309 |
-
[device to use for inference, option: "cuda" or "cpu"], by default "cuda"
|
310 |
-
nms_threshold : float, optional
|
311 |
-
[Non-maximum suppression threshold to use for the model inference, values between 0 to 1], by default 0.1
|
312 |
-
confidence_threshold : float, optional
|
313 |
-
[Confidence threshold to use for the model inference, values between 0 to 1], by default 0.7
|
314 |
-
iou_threshold : float, optional
|
315 |
-
[IOU threshold to use for identifying true positives from the predictions and ground truth], by default 0.4
|
316 |
-
gt_statistics : bool, optional
|
317 |
-
[option to generate the df_gt_analysis], by default True
|
318 |
-
|
319 |
-
Returns
|
320 |
-
-------
|
321 |
-
df_full : [pandas.DataFrame]
|
322 |
-
[dataframe containing the combined inference and ground truth information of the test images by image]
|
323 |
-
df_gt_analysis : pandas.DataFrame, optional
|
324 |
-
[dataframe containing the combined inference and ground truth information of the test images by ground truth]
|
325 |
-
"""
|
326 |
-
|
327 |
-
print(f"NMS Threshold: {nms_threshold}")
|
328 |
-
print(f"Confidence Threshold: {confidence_threshold}")
|
329 |
-
print(f"IOU Threshold: {iou_threshold}")
|
330 |
-
|
331 |
-
df_gt = get_gt_from_img_folder(
|
332 |
-
csv_file, img_folder, xml_folder, names_file)
|
333 |
-
print("Successful Generation of Ground Truth Information")
|
334 |
-
df_predictions = generate_inference_from_img_folder(
|
335 |
-
csv_file, model_cfg, img_folder, ckpt_file,
|
336 |
-
nms_thresh=nms_threshold, conf_thresh=confidence_threshold, device=device)
|
337 |
-
print("Successful Generation of Inference")
|
338 |
-
|
339 |
-
df_all = pd.merge(df_gt, df_predictions, how='left', on=["image_file_path", "image_file_name"])
|
340 |
-
print("Successful Merging")
|
341 |
-
|
342 |
-
class_labels_list = []
|
343 |
-
with open(names_file) as f:
|
344 |
-
for index,line in enumerate(f):
|
345 |
-
class_labels = line.splitlines()[0]
|
346 |
-
class_labels_list.append(class_labels)
|
347 |
-
|
348 |
-
combined_info_list = []
|
349 |
-
for _,rows in df_all.iterrows():
|
350 |
-
img_file = rows["image_file_name"]
|
351 |
-
predicted_boxes = rows["predicted_boxes"]
|
352 |
-
predicted_labels = rows["predicted_labels"]
|
353 |
-
predicted_confidence = rows["predicted_confidence"]
|
354 |
-
gt_boxes = rows["gt_boxes"]
|
355 |
-
gt_labels = rows["gt_labels"]
|
356 |
-
cm_gt_list = rows["cm_gt_list"]
|
357 |
-
cm_detections_list = rows["cm_detections_list"]
|
358 |
-
|
359 |
-
if rows["number_of_predictions"] == 0:
|
360 |
-
# Ground Truth Analysis
|
361 |
-
gt_summary_list = []
|
362 |
-
gt_match_list = []
|
363 |
-
gt_match_idx_list = []
|
364 |
-
gt_match_idx_conf_list = []
|
365 |
-
gt_match_idx_bb_list = []
|
366 |
-
for idx in range(len(gt_labels)):
|
367 |
-
gt_summary = "NO"
|
368 |
-
match = ["GT", idx, "-"]
|
369 |
-
match_idx = "-"
|
370 |
-
match_bb = "-"
|
371 |
-
gt_summary_list.append(gt_summary)
|
372 |
-
gt_match_list.append(tuple(match))
|
373 |
-
gt_match_idx_list.append(match_idx)
|
374 |
-
gt_match_idx_conf_list.append(match_idx)
|
375 |
-
gt_match_idx_bb_list.append(match_bb)
|
376 |
-
|
377 |
-
combined_info = {
|
378 |
-
"image_file_name": img_file,
|
379 |
-
"number_of_predictions_conf": [],
|
380 |
-
"predicted_labels_conf": [],
|
381 |
-
"predicted_confidence_conf": [],
|
382 |
-
"num_matches": [],
|
383 |
-
"num_mismatch": [],
|
384 |
-
"labels_hit": [],
|
385 |
-
"pairs_mislabel_gt_prediction": [],
|
386 |
-
"gt_match_idx_list": gt_match_idx_list,
|
387 |
-
"gt_match_idx_conf_list": gt_match_idx_conf_list,
|
388 |
-
"gt_match_idx_bb_list": gt_match_idx_bb_list,
|
389 |
-
"prediction_match": [],
|
390 |
-
"gt_analysis": gt_summary_list,
|
391 |
-
"prediction_analysis": [],
|
392 |
-
"gt_match": gt_match_list
|
393 |
-
}
|
394 |
-
|
395 |
-
else:
|
396 |
-
|
397 |
-
# Generate Confusion Matrix with their corresponding matches
|
398 |
-
CM = ConfusionMatrix(
|
399 |
-
num_classes=len(class_labels_list)+1,
|
400 |
-
CONF_THRESHOLD = confidence_threshold,
|
401 |
-
IOU_THRESHOLD = iou_threshold)
|
402 |
-
|
403 |
-
matching_boxes = CM.process_batch(
|
404 |
-
detections=np.asarray(cm_detections_list),
|
405 |
-
labels=np.asarray(cm_gt_list),
|
406 |
-
return_matches=True)
|
407 |
-
|
408 |
-
predicted_confidence_count = len([confidence for confidence in predicted_confidence if confidence > confidence_threshold])
|
409 |
-
predicted_confidence_round = [round(confidence, 4) for confidence in predicted_confidence]
|
410 |
-
|
411 |
-
predicted_confidence_conf = predicted_confidence_round[:predicted_confidence_count]
|
412 |
-
predicted_labels_conf = predicted_labels[:predicted_confidence_count]
|
413 |
-
predicted_boxes_conf = predicted_boxes[:predicted_confidence_count]
|
414 |
-
|
415 |
-
number_of_predictions_conf = len(predicted_labels_conf)
|
416 |
-
|
417 |
-
match_correct_list = []
|
418 |
-
match_wrong_list = []
|
419 |
-
gt_matched_idx_dict = {}
|
420 |
-
predicted_matched_idx_dict = {}
|
421 |
-
gt_mismatch_idx_dict = {}
|
422 |
-
predicted_mismatch_idx_dict = {}
|
423 |
-
labels_hit = []
|
424 |
-
pairs_mislabel_gt_prediction = []
|
425 |
-
|
426 |
-
for match in matching_boxes:
|
427 |
-
gt_idx = int(match[0])
|
428 |
-
predicted_idx = int(match[1])
|
429 |
-
iou = round(match[2], 4)
|
430 |
-
match = [gt_idx, predicted_idx, iou]
|
431 |
-
|
432 |
-
if gt_labels[gt_idx] == predicted_labels_conf[predicted_idx]:
|
433 |
-
match_correct_list.append(match)
|
434 |
-
gt_matched_idx_dict[gt_idx] = match
|
435 |
-
predicted_matched_idx_dict[predicted_idx] = match
|
436 |
-
labels_hit.append(gt_labels[gt_idx])
|
437 |
-
else:
|
438 |
-
match_wrong_list.append(match)
|
439 |
-
gt_mismatch_idx_dict[gt_idx] = match
|
440 |
-
predicted_mismatch_idx_dict[predicted_idx] = match
|
441 |
-
pairs_mislabel_gt_prediction.append(
|
442 |
-
[gt_labels[gt_idx],predicted_labels_conf[predicted_idx]])
|
443 |
-
|
444 |
-
# Ground Truth Analysis
|
445 |
-
gt_summary_list = []
|
446 |
-
gt_match_list = []
|
447 |
-
gt_match_idx_list = []
|
448 |
-
gt_match_idx_conf_list = []
|
449 |
-
gt_match_idx_bb_list = []
|
450 |
-
for idx in range(len(gt_labels)):
|
451 |
-
if idx in gt_matched_idx_dict.keys():
|
452 |
-
gt_summary = "MATCH"
|
453 |
-
match = gt_matched_idx_dict[idx]
|
454 |
-
match_idx = predicted_labels_conf[match[1]]
|
455 |
-
match_conf = predicted_confidence_conf[match[1]]
|
456 |
-
match_bb = predicted_boxes_conf[match[1]]
|
457 |
-
elif idx in gt_mismatch_idx_dict.keys():
|
458 |
-
gt_summary = "MISMATCH"
|
459 |
-
match = gt_mismatch_idx_dict[idx]
|
460 |
-
match_idx = predicted_labels_conf[match[1]]
|
461 |
-
match_conf = predicted_confidence_conf[match[1]]
|
462 |
-
match_bb = predicted_boxes_conf[match[1]]
|
463 |
-
else:
|
464 |
-
gt_summary = "NO"
|
465 |
-
match = ["GT", idx, "-"]
|
466 |
-
match_idx = "-"
|
467 |
-
match_conf = "-"
|
468 |
-
match_bb = "-"
|
469 |
-
gt_summary_list.append(gt_summary)
|
470 |
-
gt_match_list.append(tuple(match))
|
471 |
-
gt_match_idx_list.append(match_idx)
|
472 |
-
gt_match_idx_conf_list.append(match_conf)
|
473 |
-
gt_match_idx_bb_list.append(match_bb)
|
474 |
-
|
475 |
-
# Prediction Analysis
|
476 |
-
prediction_summary_list = []
|
477 |
-
prediction_match_list = []
|
478 |
-
for idx in range(len(predicted_labels_conf)):
|
479 |
-
if idx in predicted_matched_idx_dict.keys():
|
480 |
-
prediction_summary = "MATCH"
|
481 |
-
match = predicted_matched_idx_dict[idx]
|
482 |
-
elif idx in predicted_mismatch_idx_dict.keys():
|
483 |
-
prediction_summary = "MISMATCH"
|
484 |
-
match = predicted_mismatch_idx_dict[idx]
|
485 |
-
else:
|
486 |
-
prediction_summary = "NO"
|
487 |
-
match = [idx, "P", "-"]
|
488 |
-
prediction_summary_list.append(prediction_summary)
|
489 |
-
prediction_match_list.append(tuple(match))
|
490 |
-
|
491 |
-
combined_info = {
|
492 |
-
"image_file_name": img_file,
|
493 |
-
"number_of_predictions_conf": number_of_predictions_conf,
|
494 |
-
"predicted_labels_conf": predicted_labels_conf,
|
495 |
-
"predicted_confidence_conf": predicted_confidence_conf,
|
496 |
-
"num_matches": len(match_correct_list),
|
497 |
-
"num_mismatch": len(match_wrong_list),
|
498 |
-
"labels_hit": labels_hit,
|
499 |
-
"pairs_mislabel_gt_prediction": pairs_mislabel_gt_prediction,
|
500 |
-
"gt_match_idx_list": gt_match_idx_list,
|
501 |
-
"gt_match_idx_conf_list": gt_match_idx_conf_list,
|
502 |
-
"gt_match_idx_bb_list": gt_match_idx_bb_list,
|
503 |
-
"gt_match": gt_match_list,
|
504 |
-
"prediction_match": prediction_match_list,
|
505 |
-
"gt_analysis": gt_summary_list,
|
506 |
-
"prediction_analysis": prediction_summary_list
|
507 |
-
}
|
508 |
-
|
509 |
-
combined_info_list.append(combined_info)
|
510 |
-
|
511 |
-
df_combined = pd.DataFrame(combined_info_list)
|
512 |
-
|
513 |
-
df_full = pd.merge(df_all, df_combined , how='left', on=["image_file_name"])
|
514 |
-
|
515 |
-
csv_path_combined = f"{csv_save_folder}df_inference_details_nms_{nms_threshold}_conf_{confidence_threshold}_iou_{iou_threshold}.csv"
|
516 |
-
|
517 |
-
df_full.to_csv(csv_path_combined, index=False)
|
518 |
-
print ("Dataframe saved as csv to " + csv_path_combined)
|
519 |
-
|
520 |
-
if gt_statistics:
|
521 |
-
print("Generating Statistics for Single Ground Truth")
|
522 |
-
csv_path_gt = f"{csv_save_folder}df_gt_details_nms_{nms_threshold}_conf_{confidence_threshold}_iou_{iou_threshold}.csv"
|
523 |
-
df_gt_analysis = __get_single_gt_analysis(csv_output=csv_path_gt, df_input=df_full)
|
524 |
-
|
525 |
-
return df_full, df_gt_analysis
|
526 |
-
|
527 |
-
else:
|
528 |
-
return df_full
|
529 |
-
|
530 |
-
def __get_single_gt_analysis(csv_output, df_input=None,csv_input=None):
|
531 |
-
|
532 |
-
if df_input is None:
|
533 |
-
df_gt = pd.read_csv(csv_input)
|
534 |
-
|
535 |
-
# Apply literal eval of columns containing information on Ground Truth
|
536 |
-
df_gt.gt_labels = df_gt.gt_labels.apply(literal_eval)
|
537 |
-
df_gt.gt_boxes = df_gt.gt_boxes.apply(literal_eval)
|
538 |
-
df_gt.gt_boxes_int = df_gt.gt_boxes_int.apply(literal_eval)
|
539 |
-
df_gt.gt_area_list = df_gt.gt_area_list.apply(literal_eval)
|
540 |
-
df_gt.gt_area_type = df_gt.gt_area_type.apply(literal_eval)
|
541 |
-
df_gt.truncated_list = df_gt.truncated_list.apply(literal_eval)
|
542 |
-
df_gt.occluded_list = df_gt.occluded_list.apply(literal_eval)
|
543 |
-
df_gt.gt_match_idx_list = df_gt.gt_match_idx_list.apply(literal_eval)
|
544 |
-
df_gt.gt_match_idx_conf_list = df_gt.gt_match_idx_conf_list.apply(literal_eval)
|
545 |
-
df_gt.gt_match_idx_bb_list = df_gt.gt_match_idx_bb_list.apply(literal_eval)
|
546 |
-
df_gt.gt_match = df_gt.gt_match.apply(literal_eval)
|
547 |
-
df_gt.gt_analysis = df_gt.gt_analysis.apply(literal_eval)
|
548 |
-
|
549 |
-
else:
|
550 |
-
df_gt = df_input
|
551 |
-
|
552 |
-
gt_info_list = []
|
553 |
-
for _,rows in df_gt.iterrows():
|
554 |
-
# print(rows["image_file_name"])
|
555 |
-
for idx in range(rows["number_of_gt"]):
|
556 |
-
df_gt_image_dict = {
|
557 |
-
"GT_Image": rows["image_file_name"],
|
558 |
-
"GT_Label": rows["gt_labels"][idx],
|
559 |
-
"GT_Boxes": rows["gt_boxes"][idx],
|
560 |
-
"GT_Boxes_Int": rows["gt_boxes_int"][idx],
|
561 |
-
"GT_Area": rows["gt_area_list"][idx],
|
562 |
-
"GT_Area_Type": rows["gt_area_type"][idx],
|
563 |
-
"Truncated": rows["truncated_list"][idx],
|
564 |
-
"Occluded": rows["occluded_list"][idx],
|
565 |
-
"GT_Match": rows["gt_match"][idx],
|
566 |
-
"IOU": rows["gt_match"][idx][2],
|
567 |
-
"GT_Match_IDX": rows["gt_match_idx_list"][idx],
|
568 |
-
"GT_Confidence_IDX": rows["gt_match_idx_conf_list"][idx],
|
569 |
-
"GT_Predicted_Boxes_IDX": rows["gt_match_idx_bb_list"][idx],
|
570 |
-
"GT_Analysis": rows["gt_analysis"][idx]
|
571 |
-
}
|
572 |
-
gt_info_list.append(df_gt_image_dict)
|
573 |
-
|
574 |
-
df_final = pd.DataFrame(gt_info_list)
|
575 |
-
df_final = df_final.reset_index(drop=True)
|
576 |
-
|
577 |
-
df_final.to_csv(csv_output, index=False)
|
578 |
-
print ("Dataframe saved as csv to " + csv_output)
|
579 |
-
|
580 |
-
return df_final
|
581 |
-
|
582 |
-
if __name__ == '__main__':
|
583 |
-
|
584 |
-
combine_gt_predictions(
|
585 |
-
csv_file="/polyaxon-data/workspace/stee/voc_image_annotations_batch123.csv",
|
586 |
-
img_folder="/polyaxon-data/workspace/stee/data_batch123",
|
587 |
-
xml_folder="/polyaxon-data/workspace/stee/data_batch123/Annotations/",
|
588 |
-
names_file="/polyaxon-data/workspace/stee/data_batch123/obj.names",
|
589 |
-
model_cfg="cfg/cfg_frcn.ini",
|
590 |
-
ckpt_file="/polyaxon-data/workspace/stee/andy/epoch=99-step=61899.ckpt",
|
591 |
-
csv_save_folder="/polyaxon-data/workspace/stee/andy/generation/",
|
592 |
-
nms_threshold=0.9,
|
593 |
-
confidence_threshold=0.3,
|
594 |
-
iou_threshold=0.4,
|
595 |
-
gt_statistics=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/st_image_tools.py
CHANGED
@@ -279,7 +279,7 @@ class ImageTool:
|
|
279 |
masked_img = np.where(pred_mask[...,None], colour, img_pred)
|
280 |
masked_img = masked_img.astype(np.uint8)
|
281 |
|
282 |
-
img_pred = cv2.addWeighted(img_pred, 0.
|
283 |
|
284 |
def put_text_ina_mask(output, img):
|
285 |
|
@@ -319,7 +319,7 @@ class ImageTool:
|
|
319 |
|
320 |
return img
|
321 |
|
322 |
-
img_gt = cv2.addWeighted(img_gt, 0.
|
323 |
|
324 |
for output in gt_outputs:
|
325 |
img_gt = put_text_ina_mask(output, img_gt)
|
|
|
279 |
masked_img = np.where(pred_mask[...,None], colour, img_pred)
|
280 |
masked_img = masked_img.astype(np.uint8)
|
281 |
|
282 |
+
img_pred = cv2.addWeighted(img_pred, 0.7, masked_img, 0.3, 0)
|
283 |
|
284 |
def put_text_ina_mask(output, img):
|
285 |
|
|
|
319 |
|
320 |
return img
|
321 |
|
322 |
+
img_gt = cv2.addWeighted(img_gt, 0.7, masked_img, 0.3,0)
|
323 |
|
324 |
for output in gt_outputs:
|
325 |
img_gt = put_text_ina_mask(output, img_gt)
|