Add files via upload

DUE-MVSNet/losses.py

@@ -0,0 +1,38 @@
+import torch
+from torch import nn
+
+class SL1Loss(nn.Module):
+    def __init__(self, levels=3):
+        super(SL1Loss, self).__init__()
+        self.levels = levels
+        self.loss = nn.SmoothL1Loss(reduction='mean')
+
+    def forward(self, inputs, targets, masks):
+        loss = 0
+        for l in range(self.levels):
+            depth_pred_l = inputs[f'depth_{l}']
+            depth_gt_l = targets[f'level_{l}']
+            mask_l = masks[f'level_{l}']
+            if l == 1:
+                uncert_est = inputs[f'uncert_{l}']
+                disp_loss = torch.abs(depth_pred_l[mask_l] - depth_gt_l[mask_l])
+                uncert_loss = torch.exp(uncert_est[mask_l])
+                mdist_loss = disp_loss / uncert_loss
+                log_s = uncert_est[mask_l]
+                loss += mdist_loss.mean()
+                loss += log_s.mean()
+                uc_loss = log_s.mean() + mdist_loss.mean()
+            if l == 2:
+                uncert_est = inputs[f'uncert_{l}']
+                disp_loss = torch.abs(depth_pred_l[mask_l] - depth_gt_l[mask_l])
+                uncert_loss = torch.exp(uncert_est[mask_l])
+                mdist_loss = disp_loss / uncert_loss
+                log_s = uncert_est[mask_l]
+                loss += mdist_loss.mean()*0.5
+                loss += log_s.mean()*0.5
+                uc_loss = log_s.mean() + mdist_loss.mean() * 0.5
+
+            loss += self.loss(depth_pred_l[mask_l], depth_gt_l[mask_l]) * 2**(1-l)
+        return loss, uc_loss
+
+loss_dict = {'sl1': SL1Loss}