Benchmark files appear swapped between BioTrove and BioTrove-Train, plus a question about overlap with the training set
Hi, and thank you for releasing BioTrove.
I've been going through the metadata (not the images) and found a few things I wanted to raise, in case they're useful.
I'm a retired software engineer who volunteers on biodiversity data, so I may have misunderstood something.
Please correct me if I did.
- The Balanced and Unseen benchmark files look swapped between the two repos
The two repos ship the same files under opposite names:
| File | Contents | Matches README definition of ||---|---|---|
| BioTrove/BioTrove-benchmark/BioTrove-Unseen.csv | 2,253 species, 50 images each (112,650 total) | Balanced |
| BioTrove/BioTrove-benchmark/BioTrove-Balanced.csv | 300 species, 10 images each (3,000 total) | Unseen |
| BioTrove-Train/BioTrove-benchmark/BioTrove-Unseen.csv | 300 species, 10 images each (3,000 total) | Unseen |
| BioTrove-Train/BioTrove-benchmark/BioTrove-Balanced.csv | 2,253 species, 50 images each (112,650 total) | Balanced |
So the copies in BioTrove-Train look right, and the ones in BioTrove look like they have the two names switched.
(LifeStages is identical in both.) Separately, the README describes Unseen as ~11.9K images, but both copies of the 300-species file have 3,000.
- Most benchmark photos are also in BioTrove-Train
Matching on photo_id against the released BioTrove-Train parquet files:
- Unseen (300-species file): 2,314 of 3,000 photos are in BioTrove-Train
- Balanced (2,253-species file): 105,383 of 112,650 photos are in BioTrove-Train
It's possible these were removed before training and the released Train split just wasn't filtered, so my question is:
Were the models in the paper trained on a version of BioTrove-Train with the benchmark photos excluded?
If so, it might help to publish that exclusion list, or to note it in the README, so others can reproduce the results without the overlap.
- Some photos carry several different species labels
123,694 photo_ids in BioTrove (97,045 in BioTrove-Train) appear in more than one row with different species,
and 37,050 of these span kingdoms (for example a plant and an insect).
One photo, 22958149, has 13 different prairie plant species.
My guess is that these are iNaturalist photos attached to several observations (habitat shots, for example),
so each label may be right for its observation but not for what's in the photo.
For training a classifier that matters, since the same image is taught as several different species.
I have the full list as a CSV (photo_id, number of labels, species, kingdoms) and I'm happy to share it, or to open a PR adding it, if that's useful.
Thanks again for the dataset. Let me know if you'd like the scripts I used (DuckDB over the parquet files).