I examined why MiDaS can work well or not, depending on images.
If Midas work well, there has to be specific patterns. Also there has to be bad patterns either.
Then, in order to confirm there would be better and bad pattern for MiDaS Depth estimation. Especially in specific condition in image.
I took experiment with 3 patternps this time.
This is memo for my own sake, so I wrote this article as I like, not as formal.
bad pattern
This is bad pattern.Original image
Predicted image
Why this is bad pattern
2. Very close object are in sight and disturbing prediction. In this case, edge of car is the one.
3. The image had to focus on plotting target object for estimation mainly.
Total prediction stream
Better pattern
This is better pattern. I changed image conditions.
Changed conditions
2. Let target objects for estimation plot bigger than previous bad pattern.
Predicted image
Why this is better pattern
2. Cropping not moving sight(unnecessary parts) and these effect made prediction better
Total prediction stream(gif)
Best pattern
This is best pattern. I changed image conditions more.Changed conditions
2. I didn't crop width to keep image balance, which means adjusting image size well-balanced enough to see target object and easy enough to predict
3. Let target objects for estimation plot big
Original image
Predicted image
Why this is best pattern
2. Image size are well-balanced so it became easy to predict.
=> in previous case, cropped too many, though target objects are big enough to predict. However image size are unbalanced for prediction.
3. The target object is big and image size is well-balanced. These 2 make prediction work better than any pattern
Total prediction stream(gif)
To make MiDaS work better, you have to adjust conditions good for prediction.
