Added comments

model/reverse_perspective.py

@@ -26,6 +26,21 @@ class PerspectiveEstimator(nn.Module):
     Input: Pre-processed, uniformly-sized image data
     Output: Perspective factor
 
+    **Note**
+    --------
+    Loss input needs to be computed from beyond the **entire** rev-perspective
+    network. Needs to therefore compute:
+    - Effective pixel of each row after transformation.
+    - Feature density (count) along row, summed over column.
+
+    Loss is computed as a variance over row feature densities. Ref. paper 3.2.
+    After all, it is reasonable to say that you see more when you look at
+    faraway places.
+
+    This do imply that **we need to obtain a reasonably good feature extractor
+    from general images before training this submodule**. Hence, for now, we
+    prob. should work on transformer first.
+
     :param input_shape: (N, C, H, W)
     :param conv_kernel_shape: Oriented as (H, W)
     :param conv_dilation: equidistance dilation factor along H, W
@@ -96,3 +111,5 @@ class PerspectiveEstimator(nn.Module):
         out = torch.exp(-out) + self.epsilon
 
         return out
+
+