No more DPI adjusting
Closes #191 (closed)
Needs testing on a real dataset because training test is failing only because optimizers are different.
Merge request reports
Activity
requested review from @mboillet
assigned to @yschneider
changed milestone to %DAN-P5: Improve training
added 5 commits
-
fd6587de...f1ebd3ca - 2 commits from branch
main - 76b7b5f7 - No more DPI adjusting
- 2de63314 - use p=1 for equal probability
- efea7323 - use cv2.INTER_AREA interpolation
Toggle commit list-
fd6587de...f1ebd3ca - 2 commits from branch
Please train another model using this branch.
The difference between DPIAdjusting and RandomScale is how both call cv2.resize. DPIAdjusting gives a float factor for both dimensions. RandomScale computes the target size using the scaling factor and gives this size through
dsize. OpenCV recomputes both ratio based on this target size. However, since the dimensions are casted to integer, the ratio is rarely exactly equal to the scaling factor. That might explain the small difference (< 0.4%) in the images. Everything else is the same.In the last commit, I specified the
INTER_AREAinterpolation method (more suited for shrinking) to see if performances improve.added 7 commits
Toggle commit list
