def annotate():
pingit = Detector()
data ={}
data['results'] = []
rgb_folder = "/Data2TB/correctly_registered/S12/test/color/"
d_folder = "/Data2TB/correctly_registered/S12/test/depth_normalized/"
rgb_images = [f for f in listdir(rgb_folder) if isfile(join(rgb_folder, f))]
jsonprop = "/Data2TB/correctly_registered/S12/test/output.json"
jsongt = "/Data2TB/sample/annotations.json"
for image in rgb_images:
rgb_img = cv2.imread(rgb_folder + image)
d_img = cv2.imread(d_folder + image.replace(".jpg",".png"),cv2.IMREAD_GRAYSCALE)
proposals, gtobjects = common_functions.get_proposal_and_gt_boxes(image,jsonprop,jsongt)
for prop in proposals:
imgtemp_rgb = rgb_img.copy()
imgtemp_rgb = imgtemp_rgb[...,::-1]
imgtemp_d = d_img.copy()
imgtemp_d = imgtemp_d.reshape(imgtemp_d.shape[0],imgtemp_d.shape[1],1)
combined_image = np.concatenate((imgtemp_rgb, imgtemp_d), axis=2)
#cv2.imshow("tmp",imgtemp)
#cv2.waitKey()
xmin = prop[0]
ymin = prop[1]
xmax = prop[2]
ymax = prop[3]
xmin,ymin,xmax,ymax = common_functions.refine_bounding_box(xmin,ymin,xmax,ymax)
roi_combined_image = combined_image[ymin:ymax, xmin:xmax]
#roi_rgb = imgtemp_rgb[ymin:ymax, xmin:xmax]
width_original = roi_combined_image.shape[1]
width_diff = width_original - 48
roi_detect = cv2.resize(roi_combined_image, (48,48))
roi_detect = roi_detect.reshape(1,48,48,4)
roi_detect = (roi_detect.astype('float32') - 127.5)/128
cls_scores, reg,landmark = pingit.predict(roi_detect)
bb = reg[0]
lm = landmark[0]
bb,lm = common_functions.rescale_detections_to_full_image(bb,lm,width_original, prop)
ind = np.argmax(cls_scores[0])
if(ind == 1):
cv2.rectangle(rgb_img,(bb[0], bb[1]), (bb[2], bb[3]),(0,0,255), 1)
cv2.imshow("image",rgb_img)
cv2.waitKey()
def evaluate_3c(index):
config = singleton.configuration._instance.config
data = {}
data['results'] = []
tf.logging.set_verbosity(tf.logging.FATAL)
logging.getLogger("tensorflow").setLevel(logging.ERROR)
pingit = common_functions.Detector(index, config)
images_folder = config.rgb_folder
display_folder = config.display_folder
images = [
f for f in listdir(images_folder) if isfile(join(images_folder, f))
]
jsonprop = config.jsonprop
jsongt = config.jsongt
iou_props_total = 0
iou_detections_total = 0
total_detections = 0
total_gt = 0
total_tp = 0
total_fp = 0
total_pckh_precision = 0
master_count = 0
for image in images:
positive_props = []
detections_regressed = []
img = cv2.imread(images_folder + image)
proposals, gtobjects = common_functions.get_proposal_and_gt_boxes(
image, jsonprop, jsongt)
display_image = cv2.imread(display_folder + image)
count = 0
propcount = 0
for prop in proposals:
total_detections += len(proposals)
imgtemp = img.copy()
#cv2.imshow("tmp",imgtemp)
#cv2.waitKey()
xmin = prop[0]
ymin = prop[1]
xmax = prop[2]
ymax = prop[3]
xmin, ymin, xmax, ymax = common_functions.refine_bounding_box(
xmin, ymin, xmax, ymax)
roi = imgtemp[ymin:ymax, xmin:xmax]
width_original = roi.shape[1]
width_diff = width_original - 48
roi_detect = cv2.resize(roi, (48, 48))
roi_detect = roi_detect.reshape(1, 48, 48, 3)
roi_detect = (roi_detect.astype('float32') - 127.5) / 128
roi_detect = roi_detect[..., ::-1]
cls_scores, reg, landmark = pingit.predict(roi_detect)
bb = reg[0]
lm = landmark[0]
ind = np.argmax(cls_scores[0])
if (ind == 1):
positive_props.append(prop)
bb, lm = common_functions.rescale_detections_to_full_image(
bb, lm, width_original, prop)
detection = [bb, lm]
detections_regressed.append(detection)
ious = eval_functions.get_iou_totals_per_image(gtobjects,
positive_props,
detections_regressed,
images_folder + image)
#shoulder_precisions = eval_functions.get_shoulder_precisions(gtobjects, detections_regressed)
iou_props_total += ious[0]
iou_detections_total += ious[1]
total_gt += ious[2]
total_tp += ious[3]
total_fp += ious[4]
total_pckh_precision += ious[5]
#total_precisions += ious[5]
if (ious[0] != 0 and ious[1] != 0):
master_count += 1
av_iou_prop = iou_props_total / master_count
av_iou_detection = iou_detections_total / master_count
recall = total_tp / total_gt
miss_rate = total_fp / (total_tp + total_fp)
av_pckh_precision = total_pckh_precision / (master_count)
data['results'].append({
'at': str(index),
'recall': recall,
'miss_rate': miss_rate,
'av_iou': av_iou_detection,
'pckh': av_pckh_precision
})
with open(
config.results_out + config.mtcnn_model + config.bbox_loss +
".json", 'a') as outfile:
json.dump(data, outfile, indent=4)
return av_iou_detection, recall, miss_rate
def evaluate_4c(pingit, threshold, percentage):
tf.logging.set_verbosity(tf.logging.FATAL)
logging.getLogger("tensorflow").setLevel(logging.ERROR)
rgb_folder = "/Data2TB/correctly_registered/S12/test/color/"
d_folder = "/Data2TB/correctly_registered/S12/test/depth_normalized/"
display_folder = "/Data2TB/correctly_registered/S12/test/color/"
jsonprop = "/Data2TB/correctly_registered/S12/test/output.json"
jsongt = "/Data2TB/sample/annotations.json"
rgb_images = [
f for f in listdir(rgb_folder) if isfile(join(rgb_folder, f))
]
iou_props_total = 0
iou_detections_total = 0
total_gt = 0
total_tn = 0
total_tp = 0
total_fp = 0
total_outl = 0
total_fn = 0
total_pckh_precision = 0
master_count = 0
#for i in range(0,3):#image in rgb_images:
for image in rgb_images:
total_detections = 0
#image = rgb_images[i]
positive_props = []
detections_regressed = []
rgb_img = cv2.imread(rgb_folder + image)
d_img = cv2.imread(d_folder + image.replace(".jpg", ".png"),
cv2.IMREAD_GRAYSCALE)
proposals, gtobjects = common_functions.get_proposal_and_gt_boxes(
image, jsonprop, jsongt)
display_image = cv2.imread(display_folder + image)
count = 0
propcount = 0
for prop in proposals:
total_detections += 1
imgtemp_rgb = rgb_img.copy()
imgtemp_rgb = imgtemp_rgb[..., ::-1]
imgtemp_d = d_img.copy()
imgtemp_d = imgtemp_d.reshape(imgtemp_d.shape[0],
imgtemp_d.shape[1], 1)
combined_image = np.concatenate((imgtemp_rgb, imgtemp_d), axis=2)
#cv2.imshow("tmp",imgtemp)
#cv2.waitKey()
xmin = prop[0]
ymin = prop[1]
xmax = prop[2]
ymax = prop[3]
xmin, ymin, xmax, ymax = common_functions.refine_bounding_box(
xmin, ymin, xmax, ymax)
roi_combined_image = combined_image[ymin:ymax, xmin:xmax]
#roi_rgb = imgtemp_rgb[ymin:ymax, xmin:xmax]
width_original = roi_combined_image.shape[1]
width_diff = width_original - 48
roi_detect = cv2.resize(roi_combined_image, (48, 48))
roi_detect = roi_detect.reshape(1, 48, 48, 4)
roi_detect = (roi_detect.astype('float32') - 127.5) / 128
#roi_detect_rgb = cv2.resize(roi_rgb, (48,48))
#roi_detect_rgb = roi_detect_rgb.reshape(48,48,3)
#roi_detect_rgb = (roi_detect_rgb.astype('float32') - 127.5)/128
#roi_detect_rgb = roi_detect_rgb[...,::-1]
#
#roi_d = imgtemp_d[ymin:ymax, xmin:xmax]
#
#width_original = roi_d.shape[1]
#width_diff = width_original - 48
#
##r,g,b = roi_d.split()
#roi_detect_d = cv2.resize(roi_d, (48,48))
#roi_detect_d = roi_detect_d.reshape(48,48,1)
#roi_detect_d = (roi_detect_d.astype('float32') - 127.5)/128
##roi_detect = roi_detect[...,::-1]
#
##roi_detect = np.dstack((roi_detect_rgb,roi_detect_d))
#roi_detect = np.concatenate((roi_detect_rgb, roi_detect_d), axis=2)
#roi_detect = roi_detect.reshape(1,48,48,4)
#roi_detect = (roi_detect.astype('float32') - 127.5)/128
cls_scores, reg, landmark = pingit.predict(roi_detect)
bb = reg[0]
lm = landmark[0]
#ind = np.argmax(cls_scores[0])
#if(ind == 1):
if (cls_scores[0][1] >= threshold):
positive_props.append(prop)
bb, lm = common_functions.rescale_detections_to_full_image(
bb, lm, prop)
detection = [bb, lm]
detections_regressed.append(detection)
ious = get_iou_totals_per_image(gtobjects, positive_props,
detections_regressed,
rgb_folder + image, percentage)
#shoulder_precisions = eval_functions.get_shoulder_precisions(gtobjects, detections_regressed)
iou_props_total += ious[0]
iou_detections_total += ious[1]
total_gt += ious[2]
total_tp += ious[3]
total_fp += ious[4]
total_outl += ious[6]
total_fn += ious[7]
total_tn += total_detections - ious[3] - ious[4] - ious[6]
total_pckh_precision += ious[5]
#total_precisions += ious[5]
if (ious[0] != 0 and ious[1] != 0):
master_count += 1
av_iou_prop = iou_props_total / master_count
av_iou_detection = iou_detections_total / master_count
recall = total_tp / total_gt
miss_rate = total_fp / (total_fp + total_tn)
precision = total_tp / total_detections
av_pckh_precision = total_pckh_precision / (master_count)
return av_iou_detection, recall, miss_rate, av_pckh_precision, av_iou_prop
def evaluate_3c(pingit, threshold, percentage, images_folder):
data = {}
tf.logging.set_verbosity(tf.logging.FATAL)
logging.getLogger("tensorflow").setLevel(logging.ERROR)
#images_folder = "/Data2TB/correctly_registered/S12/test/color/"
display_folder = "/Data2TB/correctly_registered/S12/test/color/"
images = [
f for f in listdir(images_folder) if isfile(join(images_folder, f))
]
jsonprop = "/Data2TB/correctly_registered/S12/test/output.json"
jsongt = "/Data2TB/sample/annotations.json"
iou_props_total = 0
iou_detections_total = 0
total_gt = 0
total_tn = 0
total_tp = 0
total_fp = 0
total_outl = 0
total_fn = 0
total_pckh_precision = 0
master_count = 0
#for i in range(0,3):
for image in images:
total_detections = 0
#image = images[i]
positive_props = []
detections_regressed = []
img = cv2.imread(images_folder + image)
proposals, gtobjects = common_functions.get_proposal_and_gt_boxes(
image, jsonprop, jsongt)
display_image = cv2.imread(display_folder + image)
count = 0
propcount = 0
for prop in proposals:
total_detections += 1
imgtemp = img.copy()
#cv2.imshow("tmp",imgtemp)
#cv2.waitKey()
xmin = prop[0]
ymin = prop[1]
xmax = prop[2]
ymax = prop[3]
xmin, ymin, xmax, ymax = common_functions.refine_bounding_box(
xmin, ymin, xmax, ymax)
roi = imgtemp[ymin:ymax, xmin:xmax]
width_original = roi.shape[1]
width_diff = width_original - 48
roi_detect = cv2.resize(roi, (48, 48))
roi_detect = roi_detect.reshape(1, 48, 48, 3)
roi_detect = (roi_detect.astype('float32') - 127.5) / 128
roi_detect = roi_detect[..., ::-1]
cls_scores, reg, landmark = pingit.predict(roi_detect)
bb = reg[0]
lm = landmark[0]
if (cls_scores[0][1] >= threshold):
positive_props.append(prop)
bb, lm = common_functions.rescale_detections_to_full_image(
bb, lm, prop)
detection = [bb, lm]
detections_regressed.append(detection)
ious = get_iou_totals_per_image(gtobjects, positive_props,
detections_regressed,
images_folder + image, percentage)
#shoulder_precisions = eval_functions.get_shoulder_precisions(gtobjects, detections_regressed)
iou_props_total += ious[0]
iou_detections_total += ious[1]
total_gt += ious[2]
total_tp += ious[3]
total_fp += ious[4]
total_outl += ious[6]
total_fn += ious[7]
total_tn += total_detections - ious[3] - ious[4] - ious[6]
total_pckh_precision += ious[5]
#total_precisions += ious[5]
if (ious[0] != 0 and ious[1] != 0):
master_count += 1
av_iou_prop = iou_props_total / master_count
av_iou_detection = iou_detections_total / master_count
recall = total_tp / total_gt
miss_rate = total_fp / (total_fp + total_tn)
precision = total_tp / (total_tp + total_fp)
av_pckh_precision = total_pckh_precision / (master_count)
return av_iou_detection, recall, miss_rate, av_pckh_precision, av_iou_prop
def evaluate_4c(pingit,ayre,images_folder):
rgb_folder = "/Data2TB/correctly_registered/S12/test/color/"
images_folder = rgb_folder
d_folder = "/Data2TB/correctly_registered/S12/test/depth_normalized/"
display_folder = "/Data2TB/correctly_registered/S12/test/color/"
jsonprop = "/Data2TB/correctly_registered/S12/test/output.json"
jsongt = "/Data2TB/sample/annotations.json"
rgb_images = [f for f in listdir(rgb_folder) if isfile(join(rgb_folder, f))]
jsonprop = "/Data2TB/correctly_registered/S12/test/output.json"
jsongt = "/Data2TB/sample/annotations.json"
#list_thresholds = [0.05,0.1,0.15,0.2,0.25,0.3,0.35,0.4,0.45,0.5,0.6,0.7,0.8,0.9]
list_thresholds = [0.9]
list_percentages = [0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9]
percentage = 0.5
list_precisions = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
list_recalls = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
list_gt = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
total_gt = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
total_tn = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
total_tp = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
total_fp = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
total_outl = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
total_fn = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
total_pckh_precision = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
total_true_detections_p5 = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
iou_props_total = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
iou_detections_total = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
master_count = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
av_iou_prop = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
av_iou_detection = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
recall = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
miss_rate = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
precision = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
av_pckh_precision = [0,0,0,0,0,0,0,0,0,0,0,0,0,0]
for image in rgb_images:
total_proposals = 0
#image = images[i]
positive_props = []
detections_regressed = []
rgb_img = cv2.imread(rgb_folder + image)
d_img = cv2.imread(d_folder + image.replace(".jpg",".png"),cv2.IMREAD_GRAYSCALE)
proposals, gtobjects = common_functions.get_proposal_and_gt_boxes(image,jsonprop,jsongt)
count = 0
propcount = 0
list_positive_props = [[],[],[],[],[],[],[],[],[],[],[],[],[],[]]
list_detections_regressed = [[],[],[],[],[],[],[],[],[],[],[],[],[],[]]
for prop in proposals:
total_proposals += 1
imgtemp_rgb = rgb_img.copy()
imgtemp_rgb = imgtemp_rgb[...,::-1]
imgtemp_d = d_img.copy()
imgtemp_d = imgtemp_d.reshape(imgtemp_d.shape[0],imgtemp_d.shape[1],1)
combined_image = np.concatenate((imgtemp_rgb, imgtemp_d), axis=2)
#cv2.imshow("tmp",imgtemp)
#cv2.waitKey()
xmin = prop[0]
ymin = prop[1]
xmax = prop[2]
ymax = prop[3]
xmin,ymin,xmax,ymax = common_functions.refine_bounding_box(xmin,ymin,xmax,ymax)
roi_combined_image = combined_image[ymin:ymax, xmin:xmax]
#roi_rgb = imgtemp_rgb[ymin:ymax, xmin:xmax]
width_original = roi_combined_image.shape[1]
width_diff = width_original - 48
roi_detect = cv2.resize(roi_combined_image, (48,48))
roi_detect = roi_detect.reshape(1,48,48,4)
roi_detect = (roi_detect.astype('float32') - 127.5)/128
#roi_detect_rgb = cv2.resize(roi_rgb, (48,48))
#roi_detect_rgb = roi_detect_rgb.reshape(48,48,3)
#roi_detect_rgb = (roi_detect_rgb.astype('float32') - 127.5)/128
#roi_detect_rgb = roi_detect_rgb[...,::-1]
cls_scores, reg,landmark = pingit.predict(roi_detect)
bb = reg[0]
lm = landmark[0]
for i in range(len(list_percentages)):
if(cls_scores[0][1] >= 0.9):
list_positive_props[i].append(prop)
bbtmp,lmtmp = common_functions.rescale_detections_to_full_image(bb,lm, prop)
list_detections_regressed[i].append([bbtmp,lmtmp])
for i in range(len(list_percentages)):
results = get_iou_totals_per_image(gtobjects, list_positive_props[i], list_detections_regressed[i],images_folder + image,list_percentages[i])
iou_props_total[i] += results[0]
iou_detections_total[i] += results[1]
total_gt[i] += results[2]
total_tp[i] += results[3]
total_fp[i] += results[4]
total_outl[i] += results[6]
total_fn[i] += results[7]
total_tn[i] += total_proposals - results[3] - results[4] - results[6]
total_pckh_precision[i] += results[5]
total_true_detections_p5[i] += results[8]
#total_precisions += ious[5]
if(results[0] != 0 and results[1] != 0):
master_count[i] += 1
for i in range(len(list_percentages)):
av_iou_prop[i] = iou_props_total[i]/master_count[i]
av_iou_detection[i] = iou_detections_total[i]/master_count[i]
recall[i] = total_tp[i] / total_gt[i]
miss_rate[i] = total_fp[i] / (total_fp[i] + total_tn[i])
precision[i] = total_tp[i] / (total_tp[i] + total_fp[i])
av_pckh_precision[i] = total_pckh_precision[i]/(master_count[i])
#recall[i] = total_true_detections_p5[i] / total_gt[i]
#precision[i] = total_true_detections_p5[i] / (total_true_detections_p5[i] + total_fp[i])
#print("Threshold: " , round(list_thresholds[i],2))
#print("Average prop IOU: " , round(av_iou_prop[i],2))
#print("Average detection IOU: " , round(av_iou_detection[i],2))
#print("PCKH: " , round(av_pckh_precision[i],2))
#print("Recall: " , round(recall[i],2))
#print("Precision: " , round(precision[i],2))
print("Average PCKH: " , round(av_pckh_precision[i],2))
#print("")
#print("")
#print("Average prop IOU: " , round(av_iou_prop[13],2))
#print("Average detection IOU: " , round(av_iou_detection[13],2))
#print("Average PCKH: " , round(av_pckh_precision[6],2))
return av_iou_detection,recall, miss_rate,av_pckh_precision,av_iou_prop
def evaluate_3c(index):
config = singleton.configuration._instance.config
data = {}
data['results'] = []
tf.logging.set_verbosity(tf.logging.FATAL)
logging.getLogger("tensorflow").setLevel(logging.ERROR)
pingit = common_functions.Detector(index, config)
images_folder = config.rgb_folder
display_folder = config.display_folder
images = [
f for f in listdir(images_folder) if isfile(join(images_folder, f))
]
jsonprop = config.jsonprop
jsongt = config.jsongt
iou_props_total = 0
iou_detections_total = 0
total_detections = 0
total_gt = 0
total_tp = 0
total_fp = 0
total_pckh_precision = 0
master_count = 0
#val_folder = "/Data2TB/chl_data/mod/val/pos_neg_png_48/"
images = [
f for f in listdir(images_folder) if isfile(join(images_folder, f))
]
pos = 0
neg = 0
for image in images:
img = cv2.imread(images_folder + image, -1)
#img = cv2.imread(val_folder + "video_fifth_2018-04-23_CAM1_1524502510848_id_791_index_0.png")
positive_props = []
detections_regressed = []
img = cv2.imread(images_folder + image, -1)
#img = cv2.imread("/Data2TB/chl_data/mod/val/pos_png_48/video_fifth_2018-04-23_CAM1_1524502858816_id_831_index_0.png",-1)
##roi_detect = img.reshape(1, 48,48,3)
##roi_detect = (roi_detect.astype('float32') - 127.5)/128
##roi_detect = roi_detect[...,::-1]
proposals, gtobjects = common_functions.get_proposal_and_gt_boxes_mod(
image, jsonprop, jsongt)
display_image = cv2.imread(display_folder + image)
#
#
count = 0
propcount = 0
for prop in proposals:
total_detections += len(proposals)
imgtemp = img.copy()
#cv2.imshow("tmp",imgtemp)
#cv2.waitKey()
xmin = prop[0][0]
ymin = prop[0][1]
xmax = prop[0][2]
ymax = prop[0][3]
headpoint = prop[1]
xmin, ymin, xmax, ymax = common_functions.refine_bounding_box(
xmin, ymin, xmax, ymax)
roi = imgtemp[ymin:ymax, xmin:xmax]
x_roi = headpoint[0] - xmin
y_roi = headpoint[1] - ymin
headpoint_roi = [y_roi, x_roi]
roi = mod.get_channels(roi, headpoint_roi)
# #roi1copy = roi1.copy()
encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), 95]
result, encimg = cv2.imencode('.jpg', roi, encode_param)
roi = cv2.imdecode(encimg, 1)
#cv2.imshow("WAS IT LOS", roi)
#cv2.waitKey()
count += 1
# #continue
# #videos_holland_orbbec_2018-02-21_CAM2_1519232343120_id_66320_index_0.png
width_original = roi.shape[1]
width_diff = width_original - 48
#
roi_detect = cv2.resize(roi, (48, 48))
roi_detect = roi_detect.reshape(1, 48, 48, 3)
roi_detect = (roi_detect.astype('float32') - 127.5) / 128
roi_detect = roi_detect[..., ::-1]
#cls_scores = common_functions.testpredict(roi_detect)
cls_scores = pingit.predict(roi_detect)[0]
#print(cls_scores)
ind = np.argmax(cls_scores)
if (ind == 1):
zabri = 1
pos += 1
#cv2.imshow("pos", img)
#cv2.waitKey()
positive_props.append(prop)
#bb,lm = common_functions.rescale_detections_to_full_image(bb,lm, prop)
#detection = [bb,lm]
#detections_regressed.append(detection)
ious = eval_functions.get_iou_totals_per_image(gtobjects,
positive_props,
images_folder + image)
##shoulder_precisions = eval_functions.get_shoulder_precisions(gtobjects, detections_regressed)
#
iou_props_total += ious[0]
iou_detections_total += ious[1]
total_gt += ious[0]
total_tp += ious[1]
total_fp += ious[2]
if (ious[0] != 0 and ious[1] != 0):
master_count += 1
recall = total_tp / total_gt
miss_rate = total_fp / (total_tp + total_fp)
#av_pckh_precision = total_pckh_precision/(master_count)
data['results'].append({
'at': str(index),
'recall': recall,
'miss_rate': miss_rate
})
with open(
config.results_out + config.mtcnn_model + config.bbox_loss +
".json", 'a') as outfile:
json.dump(data, outfile, indent=4)
return recall, miss_rate
def evaluate_3c(pingit, images_folder):
images_folder = "/Data2TB/correctly_registered/S12/test/depth/"
images = [
f for f in listdir(images_folder) if isfile(join(images_folder, f))
]
jsonprop = "/Data2TB/correctly_registered/S12/test/output.json"
jsongt = "/Data2TB/sample/annotations.json"
list_thresholds = [
0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.6, 0.7, 0.8,
0.9
]
percentage = 0.5
list_precisions = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
list_recalls = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
list_gt = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
total_gt = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
total_tn = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
total_tp = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
total_fp = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
total_outl = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
total_fn = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
total_pckh_precision = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
total_true_detections_p5 = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
iou_props_total = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
iou_detections_total = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
master_count = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
av_iou_prop = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
av_iou_detection = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
recall = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
miss_rate = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
precision = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
av_pckh_precision = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
for image in images:
total_proposals = 0
#image = images[i]
positive_props = []
detections_regressed = []
img = cv2.imread(images_folder + image, -1)
proposals, gtobjects = common_functions.get_proposal_and_gt_boxes_mod(
image, jsonprop, jsongt)
count = 0
propcount = 0
list_positive_props = [[], [], [], [], [], [], [], [], [], [], [], [],
[], []]
list_detections_regressed = [[], [], [], [], [], [], [], [], [], [],
[], [], [], []]
for prop in proposals:
total_proposals += 1
imgtemp = img.copy()
#cv2.imshow("tmp",imgtemp)
#cv2.waitKey()
xmin = prop[0][0]
ymin = prop[0][1]
xmax = prop[0][2]
ymax = prop[0][3]
headpoint = prop[1]
xmin, ymin, xmax, ymax = common_functions.refine_bounding_box(
xmin, ymin, xmax, ymax)
roi = imgtemp[ymin:ymax, xmin:xmax]
x_roi = headpoint[0] - xmin
y_roi = headpoint[1] - ymin
headpoint_roi = [y_roi, x_roi]
roi = mod.get_channels(roi, headpoint_roi)
# #roi1copy = roi1.copy()
encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), 95]
result, encimg = cv2.imencode('.jpg', roi, encode_param)
roi = cv2.imdecode(encimg, 1)
width_original = roi.shape[1]
roi_detect = cv2.resize(roi, (48, 48))
roi_detect = roi_detect.reshape(1, 48, 48, 3)
roi_detect = (roi_detect.astype('float32') - 127.5) / 128
roi_detect = roi_detect[..., ::-1]
cls_scores = pingit.predict(roi_detect)[0]
for i in range(len(list_thresholds)):
if (cls_scores[0][1] >= list_thresholds[i]):
list_positive_props[i].append(prop)
#bbtmp,lmtmp = common_functions.rescale_detections_to_full_image(bb,lm, prop)
#list_detections_regressed[i].append([bbtmp,lmtmp])
for i in range(len(list_thresholds)):
results = get_iou_totals_per_image(gtobjects,
list_positive_props[i],
images_folder + image,
percentage)
iou_props_total[i] += results[0]
iou_detections_total[i] += results[1]
total_gt[i] += results[2]
total_tp[i] += results[3]
total_fp[i] += results[4]
total_outl[i] += results[6]
total_fn[i] += results[7]
total_tn[
i] += total_proposals - results[3] - results[4] - results[6]
total_pckh_precision[i] += results[5]
total_true_detections_p5[i] += results[8]
#total_precisions += ious[5]
if (results[0] != 0):
master_count[i] += 1
for i in range(len(list_thresholds)):
av_iou_prop[i] = iou_props_total[i] / master_count[i]
av_iou_detection[i] = iou_detections_total[i] / master_count[i]
recall[i] = total_tp[i] / total_gt[i]
miss_rate[i] = total_fp[i] / (total_fp[i] + total_tn[i])
precision[i] = total_tp[i] / (total_tp[i] + total_fp[i])
av_pckh_precision[i] = total_pckh_precision[i] / (master_count[i])
#recall[i] = total_tp[i] / total_gt[i]
#precision[i] = total_tp[i] / (total_tp[i] + total_fp[i])
print("Threshold: ", round(list_thresholds[i], 2))
#print("Average prop IOU: " , round(av_iou_prop[i],2))
#print("Average detection IOU: " , round(av_iou_detection[i],2))
#print("PCKH: " , round(av_pckh_precision[i],2))
print("Recall: ", round(recall[i], 2))
print("Precision: ", round(precision[i], 2))
print("")
print("")
return av_iou_detection, recall, miss_rate, av_pckh_precision, av_iou_prop