Datasets:

---

imageomics

/

2018-NEON-beetles

like 0

Follow

HDR Imageomics Institute 77

@hlapp , this look good? Relevant for SST project (see also note above). We still need this resolved as well.

@hlapp , I updated the README to hopefully remove confusion about the segmented images. We still need @sarwanali 's code though.

I have the notebook for the image segmentation. How can I provide that?

Hi @sarwanali - could you please open a PR with your code (notebook is fine) on the processing repo?

These beetles are from a mix of genera, as follows:

1. Calathus advena
2. Cyclotrachelus torvus
3. Chlaenius aestivus
4. Euryderus grossus
5. Harpalus pensylvanicus OR Pterostichus pensylvanicus --- this species epithet seems to have two different genera associated to it and will thus require more effort for alignment. I'm hoping it's just a matter of merging on an ID from the original dataset but I need to look at the filename generation to figure it out.
6. Pterostichus melanarius
7. Pterostichus restrictus
8. Pterostichus melanarius melanarius — not really 12 species, as these are subspecies level for Pterostichus melanarius
9. Pterostichus lachrymosus
10. Carabus goryi
11. Cratacanthus dubius
12. Synuchus impunctatus

Looks like images in the separate_segmented_beetle_images folder are divided into subfolders based on the associated group image and then each image gets assigned a number based on the index (in @sarwanali 's notebook), then in the Separate_segmented_train_test_splits_80_20 another _# is added to the name.

I'm not sure if these can be uniquely re-aligned though, since there are 503 folders in separate_segmented_beetle_images and all of them label images by index number. For instance, the below screenshot is Chlaenius aestivus, but beetle_245_3806.png is labeled as impunctatus (it's one of 13 images labeled beetle_245_#.png).

@egrace479 @hlapp I wondering if it is worth putting a great deal more time into trying to figure out the metadata for the work done by @sarwanali at Beetlepalooza given that it has been difficult to reconstruct. @rays45 has redone the cropping of all beetles in the dataset to individual images using grounding dino with good success as part of the Hawaii course. I know we wanted everything from Beetlepalooza to be reproducible, but I am beginning to think it might be easier to start with a set of individually cropped images for Hugging Face that are replicable rather than reconstruct what @sarwanali did as we have spent a lot of time at this point on that effort. Just my two cents

@sydnerecord , the cropped images produced by @sarwanali have been used in publications, so it is important to ensure that they are identifiable within this dataset. As it stands now, I think this PR clarifies what is and is not in the dataset at this moment and includes links to the processing code. In terms of what the data has been used for (uses of which I'm aware), this should be sufficient.

@hlapp ? I'd be comfortable merging this as-is. The remaining BeetlePalooza item I'd like to complete (before applying a DOI for clearer versioning) is generating a simple metadata file for the group images so they are displayed alongside their taxonomic information, unless there's anything else you'd like fixed first @sydnerecord ? @rays45 and I discussed the plan to add his updates after we can mark the BeetlePalooza release.

I agree on both points, thanks @egrace479 !

Just to add an update on our detection algorithm, as of present it works quite well, but is not perfect. We are running into two issues : (1) multiple beetles in a detection, (2) part of beetle getting cropped out. I have asked Mridul as a co-lead of the data paper to take over the task of improving; as I will focus on finalizing the Hawaii Beetle Dataset and run some experiments. Thanks :)