def test_sub_blocks():
start = [8]
quant = [[64,8,8]]
color = ['hsv']
for s in start:
for q in quant:
print('--- # -- ')
db_desc = SubBlockDescriptor(db)
q1_desc = SubBlockDescriptor(qs1_w1)
q2_desc = SubBlockDescriptor(qs2_w1,masks=True,mask_path=mask_root)
db_desc.compute_descriptors(grid_blocks=[s,s],quantify=q,color_space=color[0])
q1_desc.compute_descriptors(grid_blocks=[s,s],quantify=q,color_space=color[0])
q2_desc.compute_descriptors(grid_blocks=[s,s],quantify=q,color_space=color[0])
# -- SEARCH -- #
q1_search = Searcher(db_desc.result,q1_desc.result)
q2_search = Searcher(db_desc.result,q2_desc.result)
q1_desc.clear_memory()
q2_desc.clear_memory()
db_desc.clear_memory()
q1_search.search(limit=3)
q2_search.search(limit=3)
# -- EVALUATION -- #
q1_eval = EvaluateDescriptors(q1_search.result,res_root+os.sep+'gt_corresps1.pkl')
q2_eval = EvaluateDescriptors(q2_search.result,res_root+os.sep+'gt_corresps2.pkl')
q1_search.clear_memory()
q2_search.clear_memory()
q1_eval.compute_mapatk(limit=1)
q2_eval.compute_mapatk(limit=1)
filename = res_root+os.sep+'tests'+os.sep+'sub_res_'+str(s)+'_'+str(q[0])+'.pkl'
with open(filename,'wb') as f:
pickle.dump(q1_eval.score,f)
pickle.dump(q2_eval.score,f)
print('--- # -- ')
def test_level_desc():
level = [2]
start = [5,6]
jump = [2]
quant = [[16,8,8],[24,12,12]]
color = ['hsv']
for l in level:
for s in start:
for j in jump:
for q in quant:
print('--- # -- ')
db_desc = LevelDescriptor(db)
q1_desc = LevelDescriptor(qs1_w1)
q2_desc = LevelDescriptor(qs2_w1,masks=True,mask_path=mask_root)
db_desc.compute_descriptors(levels=l,init_quant=q,start=s,jump=j,color_space=color[0])
q1_desc.compute_descriptors(levels=l,init_quant=q,start=s,jump=j,color_space=color[0])
q2_desc.compute_descriptors(levels=l,init_quant=q,start=s,jump=j,color_space=color[0])
# -- SEARCH -- #
q1_search = Searcher(db_desc.result,q1_desc.result)
q2_search = Searcher(db_desc.result,q2_desc.result)
db_desc.clear_memory()
q1_desc.clear_memory()
q2_desc.clear_memory()
q1_search.search(limit=3)
q2_search.search(limit=3)
# -- EVALUATION -- #
q1_eval = EvaluateDescriptors(q1_search.result,res_root+os.sep+'gt_corresps1.pkl')
q2_eval = EvaluateDescriptors(q2_search.result,res_root+os.sep+'gt_corresps2.pkl')
q1_search.clear_memory()
q2_search.clear_memory()
q1_eval.compute_mapatk(limit=1)
q2_eval.compute_mapatk(limit=1)
filename = res_root+os.sep+'tests'+os.sep+'lev_res_'+str(l)+'_'+str(s)+'_'+str(j)+'_'+str(q[0])+'.pkl'
with open(filename,'wb') as f:
pickle.dump(q1_eval.score,f)
pickle.dump(q2_eval.score,f)
print('--- # -- ')
def main(eval_=True):
global_start = time.time()
print('-- READING IMAGES --')
start = time.time()
db_paths = sorted(glob(db_path + os.sep + '*.jpg'))
qs_paths = sorted(glob(qs1_w5 + os.sep + '*.jpg'))
db_images = [[cv2.imread(path)] for path in db_paths]
qs_images = [cv2.imread(path) for path in qs_paths]
print('-- DONE: Time: ' + str(time.time() - start))
print('-- DENOISING IMAGES --')
start = time.time()
if not os.path.isfile(res_root + os.sep + 'denoised.pkl'):
denoiser = Denoiser(qs_images)
qs_denoised = denoiser.denoise()
with open(res_root + os.sep + 'denoised.pkl', 'wb') as ff:
pickle.dump(qs_denoised, ff)
else:
with open(res_root + os.sep + 'denoised.pkl', 'rb') as ff:
qs_denoised = pickle.load(ff)
print('-- DONE: Time: ' + str(time.time() - start))
print('-- DETECTING ORIENTATION --')
start = time.time()
if not (os.path.isfile(res_root + os.sep + 'angles.pkl')
and os.path.isfile(res_root + os.sep + 'rotated.pkl')
and os.path.isfile(res_root + os.sep + 'angles_real.pkl')):
orientation = Orientation(qs_denoised)
qs_angles, qs_angles_real, qs_rotated = orientation.compute_orientation(
)
with open(res_root + os.sep + 'angles.pkl', 'wb') as ff:
pickle.dump(qs_angles, ff)
with open(res_root + os.sep + 'angles_real.pkl', 'wb') as ff:
pickle.dump(qs_angles_real, ff)
with open(res_root + os.sep + 'rotated.pkl', 'wb') as ff:
pickle.dump(qs_rotated, ff)
else:
with open(res_root + os.sep + 'angles.pkl', 'rb') as ff:
qs_angles = pickle.load(ff)
with open(res_root + os.sep + 'angles_real.pkl', 'rb') as ff:
qs_angles_real = pickle.load(ff)
with open(res_root + os.sep + 'rotated.pkl', 'rb') as ff:
qs_rotated = pickle.load(ff)
print('-- DONE: Time: ' + str(time.time() - start))
if eval_:
print('-- EVALUATING ANGLES --')
start = time.time()
angle_evaluator = EvaluateAngles(qs_angles,
qs1_w5 + os.sep + 'angles_qsd1w5.pkl')
score = angle_evaluator.evaluate(degree_margin=1.5)
print('-- DONE: Time: ' + str(time.time() - start))
print('-- SPLITTING IMAGES --')
start = time.time()
if not (os.path.isfile(res_root + os.sep + 'splitted.pkl')
and os.path.isfile(res_root + os.sep + 'qs_displays.pkl')):
spliter = SplitImages(qs_rotated)
qs_splitted, qs_displays = spliter.get_paintings()
with open(res_root + os.sep + 'splitted.pkl', 'wb') as ff:
pickle.dump(qs_splitted, ff)
with open(res_root + os.sep + 'qs_displays.pkl', 'wb') as ff:
pickle.dump(qs_displays, ff)
else:
with open(res_root + os.sep + 'splitted.pkl', 'rb') as ff:
qs_splitted = pickle.load(ff)
with open(res_root + os.sep + 'qs_displays.pkl', 'rb') as ff:
qs_displays = pickle.load(ff)
print('-- DONE: Time: ' + str(time.time() - start))
print('-- COMPUTE FOREGROUND --')
start = time.time()
if not (os.path.isfile(res_root + os.sep + 'qs_masks_rot.pkl')
and os.path.isfile(res_root + os.sep + 'qs_bboxs_rot.pkl')):
removal = BackgroundRemoval(qs_splitted)
qs_masks_rot, qs_bboxs_rot = removal.remove_background()
with open(res_root + os.sep + 'qs_masks_rot.pkl', 'wb') as ff:
pickle.dump(qs_masks_rot, ff)
with open(res_root + os.sep + 'qs_bboxs_rot.pkl', 'wb') as ff:
pickle.dump(qs_bboxs_rot, ff)
else:
with open(res_root + os.sep + 'qs_masks_rot.pkl', 'rb') as ff:
qs_masks_rot = pickle.load(ff)
with open(res_root + os.sep + 'qs_bboxs_rot.pkl', 'rb') as ff:
qs_bboxs_rot = pickle.load(ff)
print('-- DONE: Time: ' + str(time.time() - start))
print('-- UNROTATE MASKS AND FOREGROUND BOUNDING BOXES --')
start = time.time()
if not (os.path.isfile(res_root + os.sep + 'qs_masks.pkl')
and os.path.isfile(res_root + os.sep + 'qs_bboxs.pkl')):
undo_rotation = Unrotate(qs_images)
qs_masks, qs_bboxs = undo_rotation.unrotate(qs_angles, qs_bboxs_rot,
qs_masks_rot, qs_displays)
with open(res_root + os.sep + 'qs_masks.pkl', 'wb') as ff:
pickle.dump(qs_masks, ff)
with open(res_root + os.sep + 'qs_bboxs.pkl', 'wb') as ff:
pickle.dump(qs_bboxs, ff)
else:
with open(res_root + os.sep + 'qs_masks.pkl', 'rb') as ff:
qs_masks = pickle.load(ff)
with open(res_root + os.sep + 'qs_bboxs.pkl', 'rb') as ff:
qs_bboxs = pickle.load(ff)
print('-- DONE: Time: ' + str(time.time() - start))
print('-- COMPUTE FRAMES OUTPUT PICKLE --')
start = time.time()
if not os.path.isfile(res_root + os.sep + 'frames.pkl'):
qs_frames = []
for ind, bboxs in enumerate(qs_bboxs):
qs_frames.append([])
for ind2, bbox in enumerate(bboxs):
qs_frames[-1].append([qs_angles[ind], bbox])
with open(res_root + os.sep + 'frames.pkl', 'wb') as ff:
pickle.dump(qs_frames, ff)
else:
with open(res_root + os.sep + 'frames.pkl', 'rb') as ff:
qs_frames = pickle.load(ff)
print('-- DONE: Time: ' + str(time.time() - start))
print('-- COMPUTE TEXTBOXES --')
start = time.time()
if not os.path.isfile(res_root + os.sep + 'text_masks.pkl'):
text_removal = TextDetection(qs_splitted)
text_masks = text_removal.detect()
with open(res_root + os.sep + 'text_masks.pkl', 'wb') as ff:
pickle.dump(text_masks, ff)
else:
with open(res_root + os.sep + 'text_masks.pkl', 'rb') as ff:
text_masks = pickle.load(ff)
print('-- DONE: Time: ' + str(time.time() - start))
print('-- COMPUTE DESCRIPTORS --')
start = time.time()
#db_desc = SIFTDescriptor(db_images,None,None)
#qs_desc = SIFTDescriptor(qs_splitted,mask_list=qs_masks_rot,bbox_list=text_masks)
db_desc = ORBDescriptor(db_images, None, None)
qs_desc = ORBDescriptor(qs_splitted,
mask_list=qs_masks_rot,
bbox_list=text_masks)
db_desc.compute_descriptors()
qs_desc.compute_descriptors()
print('-- DONE: Time: ' + str(time.time() - start))
print('-- COMPUTE MATCHES --')
start = time.time()
matcher = MatcherFLANN(db_desc.result, qs_desc.result, flag=True)
matcher.match(min_matches=12, match_ratio=0.65)
with open('../results/result.pkl', 'wb') as ff:
pickle.dump(matcher.result, ff)
print('-- DONE: Time: ' + str(time.time() - start))
if eval_:
print('-- EVALUATING DESCRIPTORS --')
start = time.time()
desc_evaluator = EvaluateDescriptors(
matcher.result, qs1_w5 + os.sep + 'gt_corresps.pkl')
desc_evaluator.compute_mapatk(limit=1)
print('MAP@1: [{0}]'.format(desc_evaluator.score))
desc_evaluator.compute_mapatk(limit=5)
print('MAP@5: [{0}]'.format(desc_evaluator.score))
print('-- Total time: ' + str(time.time() - global_start))
def test_qs1():
print('GET DB')
db_text = get_text()
db_images = [[cv2.imread(item)]
for item in sorted(glob(os.path.join(db, "*.jpg")))]
print('done')
# -- READ -- #
print('READ FILES')
with open(res_root + os.sep + 'denoised.pkl', 'rb') as ff:
qs1_images = pickle.load(ff)
with open(res_root + os.sep + 'qs1_bbox.pkl', 'rb') as ff:
qs1_bbox = pickle.load(ff)
with open(res_root + os.sep + 'qs1_mask.pkl', 'rb') as ff:
qs1_mask = pickle.load(ff)
print('done')
# -- TRANSFORM -- #
print('computing combined descriptors')
db_desc_trans = TransformDescriptor(db_images, None, None)
db_desc_trans.compute_descriptors(transform_type='lbp')
qs_desc_trans = TransformDescriptor(qs1_images, qs1_mask, None)
qs_desc_trans.compute_descriptors(transform_type='lbp')
# -- SEARCH -- #
qs_searcher = Searcher(db_desc_trans.result, qs_desc_trans.result)
qs_searcher.search(limit=5)
print("Done.")
qs_eval = EvaluateDescriptors(qs_searcher.result,
os.path.join(qs1_w3, 'gt_corresps.pkl'))
qs_eval.compute_mapatk(limit=1)
print('DESC MAP1: [' + str(qs_eval.score) + ']')
qs_eval.compute_mapatk(limit=5)
print('DESC MAP5: [' + str(qs_eval.score) + ']')
print('done')
# -- TEXT -- #
print('computing text descriptors')
qs_desc = TextDescriptor(qs1_images, qs1_bbox)
qs_desc.compute_descriptors()
# -- SEARCH -- #
qs_searcher = SearcherText(db_text, qs_desc.result)
qs_searcher.search(limit=5)
print("Done.")
qs_eval = EvaluateDescriptors(qs_searcher.result,
os.path.join(qs1_w3, 'gt_corresps.pkl'))
qs_eval.compute_mapatk(limit=1)
print('DESC MAP1: [' + str(qs_eval.score) + ']')
qs_eval.compute_mapatk(limit=5)
print('DESC MAP5: [' + str(qs_eval.score) + ']')
print('done')
# -- COLOR -- #
print('computing color descriptors')
db_desc_col = SubBlockDescriptor(db_images, None)
db_desc_col.compute_descriptors()
qs_desc_col = SubBlockDescriptor(qs1_images, qs1_mask)
qs_desc_col.compute_descriptors()
# -- SEARCH -- #
qs_searcher = Searcher(db_desc_col.result, qs_desc_col.result)
qs_searcher.search(limit=5)
print("Done.")
qs_eval = EvaluateDescriptors(qs_searcher.result,
os.path.join(qs1_w3, 'gt_corresps.pkl'))
qs_eval.compute_mapatk(limit=1)
print('DESC MAP1: [' + str(qs_eval.score) + ']')
qs_eval.compute_mapatk(limit=5)
print('DESC MAP5: [' + str(qs_eval.score) + ']')
print('done')
# -- SEARCH -- #
qs_searcher = SearcherCombined(db_desc_col.result,
qs_desc_col.result,
db_desc_trans.result,
qs_desc_trans.result,
db_text,
qs_desc.result,
use_text=True)
db_desc_col.clear_memory()
qs_desc_col.clear_memory()
db_desc_trans.clear_memory()
qs_desc_trans.clear_memory()
qs_searcher.search(limit=5)
print("Done.")
qs_eval = EvaluateDescriptors(qs_searcher.result,
os.path.join(qs1_w3, 'gt_corresps.pkl'))
qs_eval.compute_mapatk(limit=1)
print('DESC MAP1: [' + str(qs_eval.score) + ']')
qs_eval.compute_mapatk(limit=5)
print('DESC MAP5: [' + str(qs_eval.score) + ']')
print('done')
def main_qs1w3(evaluate=False):
print("QSD1_W3")
print("Reading Images...")
db_text = get_text()
denoiser = Denoise(qs1_w3)
db_images = [[cv2.imread(item)] for item in sorted(glob(os.path.join(db,"*.jpg")))]
print("Denoising Images...")
denoiser.median_filter(3)
qs_images = denoiser.tv_bregman(weight=0.01,max_iter=1000,eps=0.001,isotropic=True)
# cv2.imwrite(r"C:\Users\PC\Documents\Roger\Master\M1\Project\Week3\tests_folder\testdenoise.png",qs_images[0][0])
# qs_images = [[cv2.imread(item)] for item in sorted(glob(os.path.join(qs1_w3,"*.jpg")))] # No denoising
print("Done.")
print("Obtaining textbox masks for each painting...")
query_mask = []
query_bbox = []
for ind,img in enumerate(qs_images):
print(ind,"of",len(qs_images))
for paint in img:
mask, textbox = TextBoxRemoval(paint)
bbox = [textbox[0][1],textbox[0][0],textbox[1][1],textbox[1][0]]
query_mask.append([mask])
query_bbox.append([bbox])
cv2.imwrite(os.path.join(res_root,'QS1W3','{0:05d}.png'.format(ind)),mask)
# cv2.imwrite(os.path.join(tests_path,'{0:05d}_mask.png'.format(ind)),mask)
print("Done.")
# input("Stop execution...")
if evaluate:
eval_iou = EvaluateIoU(query_bbox,os.path.join(qs1_w3,"text_boxes.pkl"))
eval_iou.compute_iou()
print("Bbox masks IoU:",eval_iou.score)
# -- SAVE BBOXES -- #
print("Writing final bboxs...")
with open(os.path.join(res_root,"qs1_bbox.pkl"),'wb') as file:
pickle.dump(query_bbox,file)
print("Done.")
# -- DESCRIPTORS -- #
# -- COLOR -- #
print('computing color descriptors')
db_desc_col = SubBlockDescriptor(db_images,None)
db_desc_col.compute_descriptors()
qs_desc_col = SubBlockDescriptor(qs_images,query_mask)
qs_desc_col.compute_descriptors()
# -- SEARCH -- #
qs_searcher = Searcher(db_desc_col.result,qs_desc_col.result)
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs1_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for color descriptors ", map_at_1)
print("MAP@5 for color descriptors ", map_at_5)
print("Done.")
print("Writing color desc...")
with open(os.path.join(res_root,"qs1_color_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
# -- TRANSFORM -- #
print('Obtaining transform descriptors.')
db_desc_trans = TransformDescriptor(db_images,None,None)
db_desc_trans.compute_descriptors(transform_type='hog')
qs_desc_trans = TransformDescriptor(qs_images,query_mask,None)
qs_desc_trans.compute_descriptors(transform_type='hog')
# -- SEARCH -- #
qs_searcher = Searcher(db_desc_trans.result,qs_desc_trans.result)
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs1_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for transform descriptors ", map_at_1)
print("MAP@5 for transform descriptors ", map_at_5)
print("Done.")
print("Writing transform desc...")
with open(os.path.join(res_root,"qs1_transform_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
# -- TEXT -- #
print('computing text descriptors')
qs_desc_text = TextDescriptor(qs_images,query_bbox)
qs_desc_text.compute_descriptors()
save_text(qs_desc_text.result,'qs1')
# -- SEARCH -- #
qs_searcher = SearcherText(db_text,qs_desc_text.result)
qs_desc_text.clear_memory()
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs1_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for text descriptors with levensthein ", map_at_1)
print("MAP@5 for text descriptors with levensthein ", map_at_5)
print("Done.")
print("Writing text desc...")
with open(os.path.join(res_root,"qs1_text_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
# -- COMBINED-- #
print('computing combined descriptors without text')
# -- SEARCH -- #
qs_searcher = SearcherCombined(db_desc_col.result,qs_desc_col.result,db_desc_trans.result,qs_desc_trans.result, db_text, qs_desc_text.result, False)
db_desc_col.clear_memory()
qs_desc_col.clear_memory()
db_desc_trans.clear_memory()
qs_desc_trans.clear_memory()
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs1_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for combined descriptors without text ", map_at_1)
print("MAP@5 for combined descriptors without text ", map_at_5)
print("Done.")
print("Writing combined desc...")
with open(os.path.join(res_root,"qs1_combined_without_text_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
# -- COMBINED-- #
print('computing combined descriptors with text')
# -- SEARCH -- #
qs_searcher = SearcherCombined(db_desc_col.result,qs_desc_col.result,db_desc_trans.result,qs_desc_trans.result, db_text, qs_desc_text.result, True)
db_desc_col.clear_memory()
qs_desc_col.clear_memory()
db_desc_trans.clear_memory()
qs_desc_trans.clear_memory()
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs1_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for combined descriptors with text ", map_at_1)
print("MAP@5 for combined descriptors with text ", map_at_5)
print("Done.")
print("Writing combined desc...")
with open(os.path.join(res_root,"qs1_combined_with_text_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
def main_qs2w3(evaluate=False):
# -- GET IMAGES -- #
print("Denoising Images...")
folder_path = qs2_w3
db_text = get_text()
denoiser = Denoise(folder_path)
denoiser.median_filter(3)
qs_images = denoiser.tv_bregman(weight=0.01,max_iter=1000,eps=0.001,isotropic=True)
print("Done.")
print("Obtaining list of paintings...")
img2paintings = getListOfPaintings(qs_images,"EDGES")
db_images = []
for db_path in sorted(glob(os.path.join(db,"*.jpg"))):
db_images.append([cv2.imread(db_path)])
print("Done.")
print("Obtaining background masks for each painting...")
img2paintings_mask = []
for ind,img in enumerate(img2paintings):
print(ind,"of",len(img2paintings))
img2paintings_mask.append([])
for painting in img:
mask, mean_points = BackgroundMask4(painting)
img2paintings_mask[-1].append({"painting":painting,"mask":mask,"mean_points":mean_points})
"""UNCOMMENT LINE TO PRODUCE THE MASK TO UPLOAD TO THE SERVER"""
cv2.imwrite(os.path.join(res_root,"QS2W3","{0:05d}.png".format(ind)),np.concatenate([item["mask"] for item in img2paintings_mask[-1]],axis=1))
print("Done.")
print("Obtaining textbox masks for each painting...")
img2paintings_items = []
img2paintings_bboxs = []
for ind,img in enumerate(img2paintings_mask):
print(ind,"of",len(img2paintings_mask))
img2paintings_items.append([])
for painting_items in img:
painting_masked = painting_items["painting"][painting_items["mean_points"]["top"]:painting_items["mean_points"]["bottom"],painting_items["mean_points"]["left"]:painting_items["mean_points"]["right"],:]
mask, textbox = TextBoxRemoval(painting_masked)
bbox_mask = np.zeros(shape=(painting_items["painting"].shape[0],painting_items["painting"].shape[1]))
bbox_mask[painting_items["mean_points"]["top"]:painting_items["mean_points"]["bottom"],painting_items["mean_points"]["left"]:painting_items["mean_points"]["right"]] = mask
bbox = [textbox[0][1],textbox[0][0],textbox[1][1],textbox[1][0]]
bbox[1] = bbox[1] + painting_items["mean_points"]["top"]
bbox[3] = bbox[3] + painting_items["mean_points"]["top"]
bbox[0] = bbox[0] + painting_items["mean_points"]["left"]
bbox[2] = bbox[2] + painting_items["mean_points"]["left"]
bbox_detected = False if np.mean(mask) == 255 else True
img2paintings_items[-1].append({"fg_mask":painting_items["mask"],
"mean_points":painting_items["mean_points"],
"bbox_mask":bbox_mask,
"bbox":bbox,
"bbox_detected":bbox_detected})
print("Done.")
print("Combining masks in one picture + adapting bboxes...")
final_masks = []
img2paintings_final_mask = []
img2paintings_fg_bboxs = []
final_bboxs = []
for ind,img in enumerate(img2paintings_items):
print(ind,"of",len(img2paintings_items))
to_concatenate = []
fg_bboxs = []
bboxs = []
for ind2,painting_items in enumerate(img):
total_mask = painting_items["fg_mask"]
if painting_items["bbox_detected"]:
total_mask[painting_items["bbox"][1]:painting_items["bbox"][3],painting_items["bbox"][0]:painting_items["bbox"][2]] = 0
to_concatenate.append(total_mask)
if ind2 == 0:
bboxs.append(painting_items["bbox"])
else:
missing_size = 0
for item in to_concatenate[:-1]:
missing_size += item.shape[1]
bbox = painting_items["bbox"]
bbox[0] += missing_size
bbox[2] += missing_size
bboxs.append(bbox)
fg_bboxs.append(painting_items["mean_points"])
img2paintings_fg_bboxs.append(fg_bboxs)
img2paintings_final_mask.append(to_concatenate)
final_mask = np.concatenate(to_concatenate,axis=1)
final_masks.append(final_mask)
final_bboxs.append(bboxs)
print("Done.")
if evaluate:
eval_iou = EvaluateIoU(final_bboxs,os.path.join(qs2_w3,"text_boxes.pkl"))
eval_iou.compute_iou()
print("Bbox masks IoU:",eval_iou.score)
print("Writing final bboxs...")
with open(os.path.join(res_root,"qs2_bbox.pkl"),"wb") as f:
pickle.dump(final_bboxs,f)
print("Done.")
print("Writing final masks...")
for ind,final_mask in enumerate(final_masks):
cv2.imwrite(os.path.join(res_root,"QS2W3","{0:05d}.png".format(ind)),final_mask)
print("Done.")
print("Obtaining descriptors.")
# -- DESCRIPTORS -- #
# -- COLOR -- #
print('computing color descriptors')
db_desc_col = SubBlockDescriptor(db_images,None)
db_desc_col.compute_descriptors()
qs_desc_col = SubBlockDescriptor(img2paintings,img2paintings_final_mask)
qs_desc_col.compute_descriptors()
# -- SEARCH -- #
qs_searcher = Searcher(db_desc_col.result,qs_desc_col.result)
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs2_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for color descriptors ", map_at_1)
print("MAP@5 for color descriptors ", map_at_5)
print("Done.")
print("Writing color desc...")
with open(os.path.join(res_root,"qs2_color_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
# -- TRANSFORM -- #
print('Obtaining transform descriptors.')
db_desc_trans = TransformDescriptor(db_images,None,None)
db_desc_trans.compute_descriptors(transform_type='hog')
qs_desc_trans = TransformDescriptor(img2paintings,img2paintings_final_mask,img2paintings_fg_bboxs)
qs_desc_trans.compute_descriptors(transform_type='hog')
# -- SEARCH -- #
qs_searcher = Searcher(db_desc_trans.result,qs_desc_trans.result)
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs2_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for transform descriptors ", map_at_1)
print("MAP@5 for transform descriptors ", map_at_5)
print("Done.")
print("Writing color desc...")
with open(os.path.join(res_root,"qs2_transform_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
# -- TEXT -- #
print('computing text descriptors')
qs_desc_text = TextDescriptor(img2paintings,img2paintings_fg_bboxs)
qs_desc_text.compute_descriptors()
save_text(qs_desc_text.result,'qs2')
# -- SEARCH -- #
qs_searcher = SearcherText(db_text,qs_desc_text.result)
qs_desc_text.clear_memory()
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs2_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for text descriptors with levenshtein ", map_at_1)
print("MAP@5 for text descriptors with levenshtein", map_at_5)
print("Done.")
print("Writing text desc...")
with open(os.path.join(res_root,"qs2_text_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
# -- COMBINED-- #
print('computing combined descriptors without text')
# -- SEARCH -- #
qs_searcher = SearcherCombined(db_desc_col.result,qs_desc_col.result,db_desc_trans.result,qs_desc_trans.result, db_text, qs_desc_text.result, False)
db_desc_col.clear_memory()
qs_desc_col.clear_memory()
db_desc_trans.clear_memory()
qs_desc_trans.clear_memory()
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs2_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for combined descriptors without text ", map_at_1)
print("MAP@5 for combined descriptors without text ", map_at_5)
print("Done.")
print("Writing combined desc...")
with open(os.path.join(res_root,"qs2_combined_without_text_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
# -- COMBINED-- #
print('computing combined descriptors with text')
# -- SEARCH -- #
print(db_text)
print(qs_desc_text.result)
qs_searcher = SearcherCombined(db_desc_col.result,qs_desc_col.result,db_desc_trans.result,qs_desc_trans.result, db_text, qs_desc_text.result, True)
db_desc_col.clear_memory()
qs_desc_col.clear_memory()
db_desc_trans.clear_memory()
qs_desc_trans.clear_memory()
qs_searcher.search(limit=10)
if evaluate:
evaluator = EvaluateDescriptors(qs_searcher.result, qs2_corresps_path)
map_at_1 = evaluator.compute_mapatk(1)
map_at_5 = evaluator.compute_mapatk(5)
print("MAP@1 for combined descriptors with text ", map_at_1)
print("MAP@5 for combined descriptors with text ", map_at_5)
print("Done.")
print("Writing combined desc...")
with open(os.path.join(res_root,"qs2_combined_with_text_result.pkl"),'wb') as file:
pickle.dump(qs_searcher.result,file)
print("Done.")
def evaluate():
# -- OPEN FILES -- #
print('Opening files')
with open(res_root+os.sep+'qs1_bbox.pkl','rb') as f:
qs1_bbox = pickle.load(f)
with open(res_root+os.sep+'qs2_bbox.pkl','rb') as f:
qs2_bbox = pickle.load(f)
with open(res_root+os.sep+'qs1_result.pkl','rb') as f:
qs1_result = pickle.load(f)
with open(res_root+os.sep+'qs2_result.pkl','rb') as f:
qs2_result = pickle.load(f)
print('Done')
# -- EVALUATE BBOX QS1 -- #
print('Evaluating QS1:')
qs1_box_eval = EvaluateIoU(qs1_bbox,qs1_w2+os.sep+'text_boxes.pkl')
qs1_mask_eval = EvaluateMasks(qs1_mask,qs1_w2)
qs1_desc_eval = EvaluateDescriptors(qs1_result,qs1_w2+os.sep+'gt_corresps.pkl')
qs1_box_eval.compute_iou()
qs1_mask_eval.compute_fscore()
qs1_desc_eval.compute_mapatk(limit=1)
print('DESC MAP1: ['+str(qs1_desc_eval.score)+']')
qs1_desc_eval.compute_mapatk(limit=3)
print('DESC MAP3: ['+str(qs1_desc_eval.score)+']')
print('BBOX IOU: ['+str(qs1_box_eval.score)+']')
print('MASK FSCORE: ['+str(qs1_mask_eval.score)+']')
print('Done')
# -- EVALUATE BBOX QS2 -- #
print('Evaluating QS2:')
qs2_box_eval = EvaluateIoU(qs2_bbox,qs2_w2+os.sep+'text_boxes.pkl')
qs2_mask_eval = EvaluateMasks(qs2_mask,qs2_w2)
qs2_desc_eval = EvaluateDescriptors(qs2_result,qs2_w2+os.sep+'gt_corresps.pkl')
qs2_box_eval.compute_iou()
qs2_mask_eval.compute_fscore()
qs2_desc_eval.compute_mapatk(limit=1)
print('DESC MAP1: ['+str(qs2_desc_eval.score)+']')
qs2_desc_eval.compute_mapatk(limit=3)
print('DESC MAP3: ['+str(qs2_desc_eval.score)+']')
print('BBOX IOU: ['+str(qs2_box_eval.score)+']')
print('MASK FSCORE: ['+str(qs2_mask_eval.score)+']')
print('Done')