naturalhistorymuseum / alice Goto Github PK

Need a creative solution to combine multiple masks together when they cover the same label, at different parts. At present, the multiple masks are either treated as covering different labels, or some of them are excluded further down the line.

If an image has any blank regions they should not be taken into account with the median filter

• features may be found anywhere in the image using the ORB detection algorithm
• two obvious approaches:
  1. change the image
  2. change the detection

Rare box error needs fixing. Blue line should be closest rectangle to red mask. Error only seen in a small handful of cases.

Use something like https://github.com/matterport/Mask_RCNN to find individual labels.

A model has already been trained and can roughly locate the labels, but it is not accurate enough yet to be used for extracting individual labels.

Something is going wrong with the feature matching/warping in the align labels stage.

Currently there is still blank space at the edge of the frame where the label images have been warped after being cropped. If the warping transformation was composed with the original alignment+crop then this could be avoided in almost all cases

When the software creates strips, it creates them in the wrong orientation, fully scrambling the text.
Issue observed for multiple specimens. See example:

The software periodically misses lines of text, both in handwritten and printed labels.

• Faster optical flow method

The number of cameras used in calibration is currently fixed at 4 - ideally should make this configurable so we can use more/fewer cameras

The software periodically layers semi-transparent strips over each other, decreasing readability.

e.g. when the label is tilted/curved and large parts are not visible from certain angles, few common features are found and they are often incorrect: