Computer vision is making tremendous progress in the recent times. As far as I can see the ID model used by iNaturalist is used to identify a single object in the images. However, more frequently I guess photos uploaded to iNaturalist show evidence of multiple organisms.
A bounding-box feature would entail the ability to draw a box around an organism in an observation photo and treat it as one identifiable entity.
This would provide valuable label data for upcoming versions of the computer vision model used which can thus potentially include object detection.
It will also reduce the level of duplicates in the database, crucially the level of identical photos with different labels attached.
A post was split to a new topic: Learning more about iNat’s computer vision
@jtklein this is something we’re investigating. Up to you if you want this request to remain open.
3 Likes
Yes, and it would be great if that bounding box was somehow on a separate layer, so that while it might be visible to human eyes, it would not interfere with the AI/Vision the way drawing a circle in a basic photo editor and saving that as part of the photo would.
1 Like
I am okay with closing this, but don’t know how.
closed at OP’s request
Annotated images with bounding boxes and segmentation can improve CV model performance.
LabelImg, for example, is an open source software and there are a few more that we could also consider porting over to iNat instead of building from scratch.
Thanks for reopening this feature request. I believe this is still a relevant feature and useful to broader scientific community.
@tiwane any updates on this feature, still in the works, or scrapped?
2 Likes
Not scrapped, but complicated and not a priority while we work on revamping life lists and notifications, among other more foundational work.
4 Likes