visipedia / inat_loc

The iNaturalist Localization Dataset from "On Label Granularity and Object Localization" (ECCV 2022).

The iNatLoc500 Dataset

The iNaturalist Localization 500 (iNatLoc500) dataset is a large-scale fine-grained dataset for weakly supervised object localization (WSOL). This dataset was released as part of the paper On Label Granularity and Object Localization (ECCV 2022).

Splits

Split	# Species	# Images	Avg. # Images per Species	Image-Level Labels?	Bounding Boxes?	Purpose
train	500	138k	276	Yes	No	Classifier training
val	500	12.5k	25	Yes	Yes	Localization evaluation
test	500	12.5k	25	Yes	Yes	Localization evaluation

Each image in the val and test splits has been checked to ensure that exactly one instance of the species of interest is present and that the bounding box is accurate. Full details on the dataset construction process can be found in the paper.

Label Hierarchy

iNatLoc500 is equipped with a label hierarchy based on the biological tree of life. The levels of the label hierarchy are (from finest to coarsest): species, genus, family, order, class, phylum, kingdom. Since all of the species in iNatLoc500 are animals, the kingdom level has only one node (Animalia). The iNatLoc500 dataset can be labeled at any level of the label hierarchy. For convenience we provide metadata files for each level of the label hierarchy, as described here.

Download Instructions

Instructions for downloading the dataset can be found here.

Extras

• class_mappings: Files that identify correspondences between classes in different datasets.
• source_image_mappings: Files that link iNatLoc500 images to their sources in iNat17 and iNat21.

Reference

If you find our work useful in your research please consider citing our paper:

@inproceedings{cole2022label,
  title={On Label Granularity and Object Localization},
  author={Cole, Elijah and 
          Wilber, Kimberly and 
          Van Horn, Grant and 
          Yang, Xuan and 
          Fornoni, Marco and 
          Perona, Pietro and 
          Belongie, Serge and 
          Howard, Andrew and 
          Mac Aodha, Oisin},
  booktitle={European Conference on Computer Vision},
  year={2022},
  organization={Springer}.
}

inat_loc's Issues

Taxonomy for the hierarchies

Thanks for sharing the dataset! Can you share the taxonomy of the hierarchies for this data? Ideally to load with networkx, but if it is not in that format, that should be fine too.

Image size annotations are different from the actual images

Hi!

Thanks for sharing iNatLoc500 dataset with the perfect hierarchical information! I found it very useful and helpful as an evaluation set for detection tasks.

I am actually using iNatLoc500 dataset for my research. I found the image sizes annotation provided in the dataset is quiet different from the actual image sizes. The difference are mainly caused by the swapped width and height fields in the annotation.

For instance, the test image named "test/Actinopterygii/Cyprinus carpio/05405c328abddbcbbf7596d33b7074d7.jpg" has annotation "test/Actinopterygii/Cyprinus carpio/05405c328abddbcbbf7596d33b7074d7.jpg,800,600". But the actual image size is w=800, h=600 as shown below:

In total, I found 12,085/12,500 erros in val split, and 12,066/12,500 in test split. I can make a pull request to upload the two .JSON files that recorded the difference between the annotation and the actual image sizes, and the corrected image size annotation files for val and test split.

I'd like to ask, is this caused by some rotation in the image data? Moreoever, do the bounding boxes annotations for val and test split also have corresponding errors due to the swapping? Thanks for your answer in advance! It is very important for my evaluation to know if there are also annotation drifts in the bboxes.

Thanks,

Best regards,
Mingxuan Liu