Issue: Missing data in the mpala datasets on hugging face
Dear Doctor/Professor Jenna Kline and authors of MMLA
I am currently using your dataset from the paper MMLA: Multi-Environment, Multi-Species, Low-Altitude Drone Dataset for my research. First of all, thank you and your team for making this valuable resource publicly available.
While working with the dataset, our team noticed an issue regarding some missing data (or maybe our team have some misunderstanding):
Firstly, did session1/DJI_0035 of Wilds apply down sampling while doing train test split? in mmla_wilds dataset, your paper claim that it has 13.749 frames of African wild dog. However, in README.md of https://github.com/Imageomics/mmla, the test data has 3022 images. So that, to get that number of test images, you need to need at least 30k original images (not counting the images for the training set) because of the down sampling = 10. Besides that, The DJI_0035 has 3 folders: partition_1, partition_2 and partition_3, however, script prepare_yolo_dataset.py did not recognize them because of DJI_0035_part1 and DJI_0035_part2 config (they don't have the same folder's name)
Secondly, I have counted the number of frames through hugging-face api and compared the actual of frames on hugging-face with paper's claim:
| Session | Paper's claim | Actual frames on HF | Gap |
|---|---|---|---|
| session_1 (Giraffe) | 16,891 | 16,891 | ✅ 0 |
| session_2 (Plains zebra) | 11,165 | 11,123 | ❌ -42 |
| session_3 (Grevy's zebra) | 17,940 | 17,914 | ❌ -26 |
| session_4 (Grevy's zebra) | 33,960 | 33,938 | ❌ -22 |
| session_5 (Mixed) | 24,106 | 23,236 | ❌ -870 |
| TOTAL | 104,062 | 103,102 | -960 |
Thank you very much for your time and assistance.
Best regards
Do Minh Tri