iNat-Anim

iNat-Anim Dataset

iNat-Anim (iNaturalist + Animalia) is a multi-modal few-shot image classification benchmark. It consists of 195,605 images across 673 species from an array of animal classes.

The images are taken as a subset of the images from the iNaturalist 2021 CVPR challenge [1] and are augmented with descriptions of each species from animalia.bio, an online animal encyclopedia (gathered March 2021). The descriptions are typically short and qualitatively more pertinent with respect to visual characteristics than generic species descriptions (e.g. Wikipedia).

While the raw images are included, embeddings for all images have been precomputed with a ResNet-152 model [2], allowing for efficient classifier training. Separately, we include cropped and downsampled (128x128) copies of the images, ideal for fast data exploration.

Directory structure:

• images/[id]_[class]/: Animal images of species `class`.
• inat_anim.json: Natural language class descriptions and image metadata.
• image_embeddings_resnet-152.hdf5: Precomputed ResNet-152 [2] image embeddings for all images.

[1] Grant Van Horn and Oisin Mac Aodha. 2021. iNat Challenge 2021. https://sites.google.com/view/fgvc8/competitions/inatchallenge2021?authuser=0

[2] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In _2016 IEEE Conference on Computer Vision and Pattern Recognition_, pages 770–778.