Add preprocess.py

data/preprocess.py

@@ -0,0 +1,33 @@
+import os
+import skimage.util as util
+from skimage import io
+from skimage.transform import resize
+
+
+with open('train.txt', 'r') as fd:
+    image_files = fd.readlines()
+
+total = len(image_files)
+cnt = 0
+
+# path/to/deepfashion directory
+root = '/path/to/deepfashion'
+# path/to/save directory
+save_root = 'path/to/save'
+
+for image_file in image_files:
+    image_file = os.path.join(root, image_file).strip()
+    image = io.imread(image_file)
+    pad_width_1 = (1101-750) // 2
+    pad_width_2 = (1101-750) // 2 + 1
+    image_pad = util.pad(image, ((0,0),(pad_width_1, pad_width_2),(0,0)), constant_values=232)
+    image_resize = resize(image_pad, (1024, 1024))
+    image_resize = (image_resize * 255).astype('uint8')
+    dst_file = os.path.dirname(image_file).replace(root, save_root)
+    os.makedirs(dst_file, exist_ok=True)
+    dst_file = os.path.join(dst_file, os.path.basename(image_file))
+    # dst_file = dst_file.replace('.jpg', '.png')
+    io.imsave(dst_file, image_resize)
+    cnt += 1
+    if cnt % 20 == 0:
+        print('Processing: %d / %d' % (cnt, total))

models/networks/correspondence.py

@@ -193,7 +193,7 @@ class NoVGGHPMCorrespondence(BaseNetwork):
                 dx = dx.view(1, search_window ** 2, 1, 1) * dilation
                 dy = dy.view(1, search_window ** 2, 1, 1) * dilation
                 offset_x_up = F.interpolate((2 * offset_x + dx), scale_factor=2)
-                offset_y_up = F.interpolate((2 * offset_y + dx), scale_factor=2)
+                offset_y_up = F.interpolate((2 * offset_y + dy), scale_factor=2)
             ref = ref.view(batch_size, 3, scale * scale)
             mat, y = self.patch_match(f1, f2, ref, offset_x_up, offset_y_up)
             y = y.view(batch_size, 3, scale, scale)