Semi-Aves, Semi-Fungi, Semi-CUB, Semi-iNat

Dataset Overview

Dataset List

1. Semi‑Aves: Dataset from the Semi‑Aves Challenge used in the CVPR 2020 FGVC7 workshop.
2. Semi‑Fungi: Built from the 2018 FGVCx Fungi Classification Challenge for the CVPR 2018 FGVC5 workshop.
3. Semi‑CUB: Constructed from the Caltech‑UCSD Birds‑200‑2011 dataset.
4. Semi‑iNat: New dataset for the CVPR 2021 FGVC8 workshop Semi‑iNat Challenge.

Dataset Structure

Each dataset’s split information is stored in data/${dataset}/${split}.txt and includes:

• l_train: labeled in‑domain data
• u_train_in: unlabeled in‑domain data
• u_train_out: unlabeled out‑of‑domain data
• u_train: union of u_train_in and u_train_out
• val: validation set
• l_train_val: union of l_train and val
• test: test set

Each line lists a filename and its corresponding class label.

Storage

• Semi‑Aves: data stored in data/semi_aves.
• Semi‑Fungi and Semi‑CUB: images stored in data/semi_fungi/images and data/cub/images respectively.

Notes

• Semi‑Fungi: Images resized so the longest side is 300 px.
• Semi‑Aves: Provides additional cross‑validation splits; labels are withheld to prevent leakage of unlabeled data.
• Semi‑Aves and Semi‑Fungi: Species name files are included.

Training and Evaluation

CVPR Papers

Code is provided for supervised training, self‑training, pseudo‑labeling, and curriculum pseudo‑labeling, based on this PyTorch implementation.

BMVC Papers

Hierarchical supervision based on coarse labels is added for semi‑supervised learning. Commands and parameters for hierarchical training are supplied.

Pre‑trained Models

Supervised models, MoCo pre‑trained models, and MoCo + supervised models are available for Semi‑Aves and Semi‑Fungi. Models can be downloaded via the provided links.

Related Competitions

• Semi‑iNat 2021 Competition and Semi‑Aves 2020 Competition offer challenge websites, Kaggle links, and technical reports.