identify invalid flow with 0 instead of NaN

NaN triggered lots of differentation problems

Author: Clément Pinard

datasets/KITTI.py

@@ -21,7 +21,8 @@ def load_flow_from_png(png_path):
     invalid = (flo_file[:,:,0] == 0)
     flo_img = flo_img - 32768
     flo_img = flo_img / 64
-    flo_img[invalid, :] = float('nan')
+    flo_img[np.abs(flo_img) < 1e-10] = 1e-10
+    flo_img[invalid, :] = 0
     return(flo_img)
 
 

main.py

@@ -339,6 +339,7 @@ def __repr__(self):
 def flow2rgb(flow_map, max_value):
     global args
     _, h, w = flow_map.shape
+    flow_map[:,(flow_map[0] == 0) & (flow_map[1] == 0)] = float('nan')
     rgb_map = np.ones((h,w,3)).astype(np.float32)
     if max_value is not None:
         normalized_flow_map = flow_map / max_value

models/FlowNetS.py

@@ -1,10 +1,9 @@
 import torch
 import torch.nn as nn
 from torch.nn.init import kaiming_normal
-import math
 
 __all__ = [
-    'FlowNetS', 'flownets', 'flownets_bn'
+    'flownets', 'flownets_bn'
 ]
 
 

multiscaleloss.py

@@ -6,8 +6,8 @@ def EPE(input_flow, target_flow, sparse=False, mean=True):
     EPE_map = torch.norm(target_flow-input_flow,2,1)
     batch_size = EPE_map.size(0)
     if sparse:
-        # invalid flow is defined with one of flow coordinate to be NaN
-        mask = (target_flow[:,0] != target_flow[:,0]) | (target_flow[:,1] != target_flow[:,1])
+        # invalid flow is defined with both flow coordinates to be exactly 0
+        mask = (target_flow[:,0] == 0) & (target_flow[:,1] == 0)
 
         EPE_map = EPE_map[~mask.data]
     if mean:
@@ -16,11 +16,9 @@ def EPE(input_flow, target_flow, sparse=False, mean=True):
         return EPE_map.sum()/batch_size
 
 
-def sparse_max_pool(target, size):
-    input = target.clone()
-    input[(target != target).data] = 0
-    positive = (target > 0).float()
-    negative = (target < 0).float()
+def sparse_max_pool(input, size):
+    positive = (input > 0).float()
+    negative = (input < 0).float()
     output = nn.functional.adaptive_max_pool2d(input * positive, size) - nn.functional.adaptive_max_pool2d(-input * negative, size)
     return output