vato007/fast-depth-tf
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix total loss calculation, add some more todos

Browse Source
This commit is contained in:
Piv
2022-05-03 19:05:29 +09:30
parent 188c55d1c8
commit 909c38cf56
1 changed files with 7 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
unsupervised/train.py
View File

@@ -55,6 +55,7 @@ class UnsupervisedPoseDepthLearner(keras.Model):
# It isn't too much of an issue right now since we're only using 2 images (left/right) # It isn't too much of an issue right now since we're only using 2 images (left/right)
# For each depth output (scale), do the projective inverse warp on each input image and calculate the losses # For each depth output (scale), do the projective inverse warp on each input image and calculate the losses
# Only take the min loss between the two warped images (from monodepth2) # Only take the min loss between the two warped images (from monodepth2)
# TODO: Need to bilinear resize the depth at each scale up to the size of image
warp1 = warp.projective_inverse_warp(data.frames[0], depth[scale], pose1, data.intrinsics, coords) warp1 = warp.projective_inverse_warp(data.frames[0], depth[scale], pose1, data.intrinsics, coords)
warp2 = warp.projective_inverse_warp(data.frames[2], depth[scale], pose1, data.intrinsics, coords) warp2 = warp.projective_inverse_warp(data.frames[2], depth[scale], pose1, data.intrinsics, coords)
@@ -73,13 +74,17 @@ class UnsupervisedPoseDepthLearner(keras.Model):
tf.reduce_min(tf.concat([warp_loss1, warp_loss2, source_loss1, source_loss2], axis=3), axis=3)) tf.reduce_min(tf.concat([warp_loss1, warp_loss2, source_loss1, source_loss2], axis=3), axis=3))
# Calculate smooth losses # Calculate smooth losses
# TODO: Since smooth loss is calculated directly on the depth at the scale, we need
# to resize the target image to the same dimensions as the depth map at the current scale
# Can do this by just inspecting the shape of the depth and resizing to match that (but
# with 3 colour channels)
smooth_loss = loss.smooth_loss(depth[scale], data.frames[1]) smooth_loss = loss.smooth_loss(depth[scale], data.frames[1])
# TODO: Monodepth also divides the smooth loss by 2 ** scale, why? # SFM Learner downscales smoothing loss depending on the scale
smoothed_reprojection_loss = self.smoothness * smooth_loss / (2 ** scale) smoothed_reprojection_loss = self.smoothness * smooth_loss / (2 ** scale)
# Add to total loss (with smooth loss + smooth loss weighting applied to pixel losses) # Add to total loss (with smooth loss + smooth loss weighting applied to pixel losses)
total_loss += reprojection_loss + smooth_loss total_loss += reprojection_loss + smoothed_reprojection_loss
pass pass
# Collect losses, average them out (divide by number of scales) # Collect losses, average them out (divide by number of scales)