Update README, add poisson(remain questions) and fixes, tbd: temporal…

… filter, static mode.

README.md

@@ -1,2 +1,11 @@
 # motion_magnification_learning-based
-An unofficial implementation of "Learning-based Video Motion Magnification" in Pytorch.
+An unofficial implementation of "[Learning-based Video Motion Magnification](https://arxiv.org/abs/1804.02684)" in Pytorch.
+Refer to the official tensorflow version [code](https://github.com/12dmodel/deep_motion_mag).
+
+# Data preparation
+1. coco100000: please refer to the clarification in the official respository.
+2. frames of validation video: TBD. 
+
+# Run
+`bash run_G.sh` to train and test
+> feel free to _ctrl+c_ to stop train, it will automatically do testing on several well-prepared validation video frames.

config.py

@@ -9,7 +9,7 @@ def __init__(self, lambda_G_new=None, skip=None, videos_train=None):
         self.GPUs = '0'
         self.batch_size = 4 * len(self.GPUs.split(','))
         self.use_energy = False
-        self.date = '2019-11-24'
+        self.date = '2019-11-26'
         prefix = '..'
         self.dir_train = os.path.join(prefix, 'datasets/motion_mag_data/train')
         self.dir_test = os.path.join(prefix, 'datasets/motion_mag_data/test')
@@ -40,7 +40,7 @@ def __init__(self, lambda_G_new=None, skip=None, videos_train=None):
         self.skip = 0 if not isinstance(skip, int) else skip
         self.batch_size_test = 1
         self.video_num = len(self.videos_train)
-        self.preproc = ['resize']
+        self.preproc = ['resize', 'poisson']
 
         self.lr = 1e-4
         self.betas_G = (0.9, 0.999)

data.py

@@ -12,6 +12,17 @@
 ImageFile.LOAD_TRUNCATED_IMAGES = True
 Tensor = torch.cuda.FloatTensor
 
+
+def gen_poisson_noise(unit):
+    n = np.random.randn(*unit.shape)
+
+    # Strange here, unit has been in range of (-1, 1),
+    # but currently same as the official codes.
+    n_str = np.sqrt(unit + 1.0) / np.sqrt(127.5)
+    poisson_noise = np.multiply(n, n_str)
+    return poisson_noise
+
+
 def load_unit(path, preproc):
     # Load
     file_suffix = path.split('.')[-1].lower()
@@ -34,6 +45,8 @@ def load_unit(path, preproc):
     unit = cv2.cvtColor(unit, cv2.COLOR_BGR2RGB)
     try:
         unit = unit.astype(np.float32) / 127.5 - 1.0
+        if 'poisson' in preproc:
+            unit = unit + gen_poisson_noise(unit) * np.random.uniform(0, 0.3)
     except Exception as e:
         print('EX:', e, unit.shape, unit.dtype)
 
@@ -50,7 +63,7 @@ def unit_postprocessing(unit, residual=None, vid_size=None):
         unit = torch.clamp(unit - residual, -1.0, 1.0)
     unit = unit.cpu().detach().numpy()
     unit = np.round((np.transpose(unit, (1, 2, 0)) + 1.0) * 127.5).astype(np.uint8)
-    if unit.shape[:2][::-1] != vid_size:
+    if unit.shape[:2][::-1] != vid_size and vid_size is not None:
         unit = cv2.resize(unit, vid_size, interpolation=cv2.INTER_CUBIC)
     return unit
 
@@ -147,8 +160,8 @@ def gen_test(self, anchor=None):
             anchor = self.anchor
 
         for _ in range(self.batch_size):
-            unit_A = load_unit(self.paths[anchor], preproc=self.preproc)
-            unit_C = load_unit(self.paths[anchor].replace('frameA', 'frameC'), preproc=self.preproc)
+            unit_A = load_unit(self.paths[anchor], preproc=['resize'])
+            unit_C = load_unit(self.paths[anchor].replace('frameA', 'frameC'), preproc=['resize'])
             batch_A.append(unit_A)
             batch_C.append(unit_C)