IbProgrammmer Claude Opus 5.5 commited on
Commit
b123fa1
·
1 Parent(s): 0cb72a1

fix(enhance): normalize FFA-Net input like the official test code

Browse files

FFA-Net (official net/test.py) feeds the network RGB [0,1] input normalized
with mean (0.64, 0.6, 0.58) and std (0.14, 0.15, 0.152). Without it the
outdoor weights washed foggy images out (gray-level std 64.8 -> 14.4 on
RTTS BD_Baidu_046) and RTMDet found nothing. With it the image keeps its
contrast (std 67.0) and edges sharpen.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Files changed (1) hide show
  1. app.py +7 -1
app.py CHANGED
@@ -331,7 +331,13 @@ def _load_ffa(weights_path: str):
331
  self.post = nn.Sequential(conv(self.dim, self.dim, k),
332
  conv(self.dim, 3, k))
333
 
334
- def forward(self, x1):
 
 
 
 
 
 
335
  x = self.pre(x1)
336
  res1 = self.g1(x)
337
  res2 = self.g2(res1)
 
331
  self.post = nn.Sequential(conv(self.dim, self.dim, k),
332
  conv(self.dim, 3, k))
333
 
334
+ # Official FFA-Net test.py normalizes the RGB [0,1] input with these
335
+ # statistics; without it the outdoor weights wash the image out.
336
+ MEAN = torch.tensor([0.64, 0.6, 0.58]).view(1, 3, 1, 1)
337
+ STD = torch.tensor([0.14, 0.15, 0.152]).view(1, 3, 1, 1)
338
+
339
+ def forward(self, x):
340
+ x1 = (x - self.MEAN.to(x.device)) / self.STD.to(x.device)
341
  x = self.pre(x1)
342
  res1 = self.g1(x)
343
  res2 = self.g2(res1)