diff --git a/src/SumMeVideo.py b/src/SumMeVideo.py
index d86f5c16a6a80b2df3402e794d15d3c61fac5035..3857203f5d54e060b2d40180a371ce9ea29c212a 100644
--- a/src/SumMeVideo.py
+++ b/src/SumMeVideo.py
@@ -1,6 +1,7 @@
 from scipy.io import loadmat
 import os
 from moviepy.editor import VideoFileClip
+import numpy as np
 
 
 class SumMeVideo(VideoFileClip):
@@ -13,7 +14,21 @@ class SumMeVideo(VideoFileClip):
 
     def get_gt(self):
         video_gt = loadmat(self.gt_path)
-        return video_gt['gt_score']
+        video_gt = video_gt['user_score']  # (n_frames,n_annotator)
+        return self.bin_classify_user_score(video_gt)
+
+    """
+    :param user scores
+    :returns binary nparray
+    """
+
+    def bin_classify_user_score(self, video_gt):
+        user_scores = video_gt
+        result = []
+        for user in range(user_scores.shape[1]):
+            print(user_scores[:, user].shape)
+            result.append([int(item > 0.5) for item in user_scores[:, user]])
+        return np.asarray(result).T  # (n_frames,n_annotator)
 
     def get_frames(self):
         return list(self.video_clip.iter_frames(with_times=False))
diff --git a/src/TVSumVideo.py b/src/TVSumVideo.py
index 29b849df5b21d01d2f6b254b8cbefb57b0f1f305..132a9beb6c91777959a2a0e0ef070dedb4cb4b90 100644
--- a/src/TVSumVideo.py
+++ b/src/TVSumVideo.py
@@ -21,7 +21,8 @@ class TVSumVideo(VideoFileClip):
         for i in range(len(sub_gt)):
             users_gt.append(sub_gt.iloc[i, -1].split(","))
         users_gt = np.array(users_gt)
-        return self.__avg_array(users_gt)
+        avg_gt = self.__avg_array(users_gt)
+        return np.expand_dims(avg_gt, axis=1)
 
     def __avg_array(self, users_gt):
         users_gt = users_gt.astype(int)
diff --git a/src/features.pickle b/src/features.pickle
new file mode 100644
index 0000000000000000000000000000000000000000..cb2b97355b07b3ed2fbe1b642697f90930b27880
Binary files /dev/null and b/src/features.pickle differ
diff --git a/src/utils.py b/src/utils.py
index 1d00346e8e62bc1eebba514ecc7cfe22ac338335..1b9c91ce49a8ce8ed854d81d907fa0531051c4dc 100644
--- a/src/utils.py
+++ b/src/utils.py
@@ -54,27 +54,42 @@ def resize_image(img_path, width, height):
     return resized
 
 
-def numbers_to_classes(array, n_classes):
-    # TODO
-    return 0
-
-
-def sample_from_video_with_gt(video_frames, video_gt, duration, fps, n_samples=2):
+def sample_from_video_with_gt(video_frames, user_scores, duration, fps, n_samples=2):
     sampled_frames = []
     sampled_gt = []
-    source_array = np.arange(fps);
+    source_array = np.arange(fps)
     for i in range(duration):
         sorted_samples = make_sorted_sample(source_array, n_samples)
-        for idx in sorted_samples:
-            sampled_frames.append(video_frames[i * fps:(i + 1) * fps][idx - 1])
-            sampled_gt.append(video_gt[i * fps:(i + 1) * fps][idx - 1])
+        start = i * fps
+        end = (i + 1) * fps
+        time_span = (start, end)
+        sampled_frames.extend(select_frames_from_video(sorted_samples, video_frames, time_span))
+        sampled_gt.extend(select_user_scores(sorted_samples, user_scores, time_span))
     return sampled_frames, sampled_gt
 
 
+def select_frames_from_video(samples, video_frames, time_span):
+    sampled_frames = []
+    for idx in samples:
+        sampled_frames.append(video_frames[time_span[0]:time_span[1]][idx - 1])
+    return sampled_frames
+
+
+def select_user_scores(samples, user_scores, time_span):
+    sampled_gt = []
+    for idx in samples:
+        gt = [user_scores[time_span[0]:time_span[1], user][idx - 1] for user in range(user_scores.shape[1])]
+        sampled_gt.append(np.asarray(gt))
+    return sampled_gt
+
+
 def make_sorted_sample(source_array, n_samples):
     sorted_samples = np.sort(np.random.choice(source_array, n_samples, replace=False))
     return sorted_samples
 
 
 def drop_file_extension(file_name):
-    return file_name.split('.')[0]
+    if file_name is None:
+        raise ValueError
+    file_name = file_name.split('.')[:-1]
+    return '.'.join(file_name)
\ No newline at end of file
diff --git a/src/video2frames.py b/src/video2frames.py
index b456fb98c1f3b67840210668b96abd4bea8c558c..891b12358e2e9fc76fd6f906aa19c9d2bfaebd99 100644
--- a/src/video2frames.py
+++ b/src/video2frames.py
@@ -17,9 +17,9 @@ def create_video_obj(dataset, video_name, video_path, gt_dir):
 
 def arg_parser():
     parser = argparse.ArgumentParser(description='Extract Features')
-    parser.add_argument('--dataset', default='TVSum', type=str, help='SumMe or TVSum')
-    parser.add_argument('--videos_dir', metavar='DIR', default='../data/TVSum/video', help='path input videos')
-    parser.add_argument('--gt', metavar='GT_Dir', default='../data/TVSum/data', help='path ground truth')
+    parser.add_argument('--dataset', default='SumMe', type=str, help='SumMe or TVSum')
+    parser.add_argument('--videos_dir', metavar='DIR', default='../data/SumMe/videos/test', help='path input videos')
+    parser.add_argument('--gt', metavar='GT_Dir', default='../data/SumMe/GT', help='path ground truth')
     parser.add_argument('--fps', default=2, type=int, help='Frames per second for the extraction')
     parser.add_argument('--model_arch', default='googlenet',
                         help='pretrained model architecture e.g. resnet50 or alexnet')