# digit-detection models and setting
detect_model = "model/digit-detector/detector_model.hdf5"
recognize_model = "model/digit-detector/recognize_model.hdf5"
mean_value_for_detector = 107.524
mean_value_for_recognizer = 112.833
model_input_shape = (32,32,1)
preproc_for_detector = preproc.GrayImgPreprocessor(mean_value_for_detector)
preproc_for_recognizer = preproc.GrayImgPreprocessor(mean_value_for_recognizer)
char_detector = cls.CnnClassifier(detect_model, preproc_for_detector, model_input_shape)
char_recognizer = cls.CnnClassifier(recognize_model, preproc_for_recognizer, model_input_shape)
digit_spotter = detector.DigitSpotter(char_detector, char_recognizer, rp.MserRegionProposer())
# open the csv file of player-face-detection and find out the cloths color for each teams
df = pd.read_csv('C:/Users/shing/Desktop/player_info.csv', names =('Team','Num','Name','R','G','B'))
RDB_med = df.groupby('Team').median()
R0_med, G0_med, B0_med = RDB_med.iloc[0,1:4]
R1_med, G1_med, B1_med = RDB_med.iloc[1,1:4]
def process (input_image, params, model_params):
#oriImg = cv2.imread(input_image) # B,G,R order
oriImg = input_image
multiplier = [x * model_params['boxsize'] / oriImg.shape[0] for x in params['scale_search']]
heatmap_avg = np.zeros((oriImg.shape[0], oriImg.shape[1], 19))
n_files_to_sample=1000,
random_order=False)
annotator = ann.SvhnAnnotation(ANNOTATION_FILE)
preprocessor_for_detector = preproc.GrayImgPreprocessor(
mean_value_for_detector)
preprocessor_for_recognizer = preproc.GrayImgPreprocessor(
mean_value_for_recognizer)
detector = cls.CnnClassifier(detect_model, preprocessor_for_detector,
model_input_shape)
recognizer = cls.CnnClassifier(recognize_model,
preprocessor_for_recognizer,
model_input_shape)
proposer = rp.MserRegionProposer()
# 2. create detector
det = detect.DigitSpotter(detector, recognizer, proposer)
# 3. Evaluate average precision
evaluator = eval.Evaluator(det, annotator, rp.OverlapCalculator())
recall, precision, f1_score = evaluator.run(img_files)
# recall value : 0.1757938334100322, precision value : 0.18070009460737937, f1_score : 0.1782132027058549
# 4. Evaluate MSER
detector = cls.TrueBinaryClassifier(input_shape=model_input_shape)
preprocessor = preproc.NonePreprocessor()
# Todo : detector, recognizer 를 none type 으로
det = detect.DigitSpotter(detector, recognizer, proposer)
POS_OVERLAP_THD = 0.6
PATCH_SIZE = (32, 32)
if __name__ == "__main__":
# 1. file 을 load
files = file_io.list_files(directory=DIR,
pattern="*.png",
recursive_option=False,
n_files_to_sample=N_IMAGES,
random_order=False)
n_files = len(files)
n_train_files = int(n_files * 0.8)
print n_train_files
extractor = extractor_.Extractor(rp.MserRegionProposer(),
ann.SvhnAnnotation(ANNOTATION_FILE),
rp.OverlapCalculator())
train_samples, train_labels = extractor.extract_patch(
files[:n_train_files], PATCH_SIZE, POS_OVERLAP_THD, NEG_OVERLAP_THD)
extractor = extractor_.Extractor(rp.MserRegionProposer(),
ann.SvhnAnnotation(ANNOTATION_FILE),
rp.OverlapCalculator())
validation_samples, validation_labels = extractor.extract_patch(
files[n_train_files:], PATCH_SIZE, POS_OVERLAP_THD, NEG_OVERLAP_THD)
print train_samples.shape, train_labels.shape
print validation_samples.shape, validation_labels.shape
# show.plot_images(samples, labels.reshape(-1,).tolist())
N_IMAGES = None
DIR = '../dataset/svhn/train'
ANNOTATION_FILE = "../dataset/svhn/train/digitStruct.json"
NEG_OVERLAP_THD = 0.05
POS_OVERLAP_THD = 0.6
PATCH_SIZE = (32,32)
if __name__ == "__main__":
# 1. file 을 load
files = file_io.list_files(directory=DIR, pattern="*.png", recursive_option=False, n_files_to_sample=N_IMAGES, random_order=False)
n_files = len(files)
n_train_files = int(n_files * 0.8)
print (n_train_files)
extractor = extractor_.Extractor(rp.MserRegionProposer(), ann.SvhnAnnotation(ANNOTATION_FILE), rp.OverlapCalculator())
train_samples, train_labels = extractor.extract_patch(files[:n_train_files], PATCH_SIZE, POS_OVERLAP_THD, NEG_OVERLAP_THD)
extractor = extractor_.Extractor(rp.MserRegionProposer(), ann.SvhnAnnotation(ANNOTATION_FILE), rp.OverlapCalculator())
validation_samples, validation_labels = extractor.extract_patch(files[n_train_files:], PATCH_SIZE, POS_OVERLAP_THD, NEG_OVERLAP_THD)
print (train_samples.shape, train_labels.shape)
print (validation_samples.shape, validation_labels.shape)
# show.plot_images(samples, labels.reshape(-1,).tolist())
file_io.FileHDF5().write(train_samples, "train.hdf5", "images", "w", dtype="uint8")
file_io.FileHDF5().write(train_labels, "train.hdf5", "labels", "a", dtype="int")
file_io.FileHDF5().write(validation_samples, "val.hdf5", "images", "w", dtype="uint8")
file_io.FileHDF5().write(validation_labels, "val.hdf5", "labels", "a", dtype="int")