Semi-Supervised iNaturalist Challenge at FGVC8

This semi-supervised recognition challenge focuses on recognizing animal, plant, and fungi species from partially labelled data. This dataset is designed to expose some of the challenges encountered in a realistic application, such as the highly similar classes, class imbalance, and domain mismatch between the labeled and unlabeled data. Different from last year's challenge, this challenge includes more species that span different kingdoms, and does not make a distinction between in-class and out-of-class unlabeled data.

This challenge is part of the FGVC8 workshop at CVPR 2021. Teams with top submissions, at the discretion of the workshop organizers, will be invited to present their work at the workshop.

The competition will be hosted on Kaggle.

Please open an issue here (or email jcsu@cs.umass.edu) if you have questions or problems with the dataset.

Updates

• 3/8/2021: Dataset released, challenge starts.
• 3/30/2021: The annotation has been updated, please use the updated version (annotation_v2.json).
• 6/1/2021: The Kaggle competition has ended. We will have a discussion of the competition at the FGVC8 workshop on 6/25/2021.
• 6/3/2021: The tech report of the dataset is now on arXiv.
• 7/23/2021: Test set labels, and the labels for unlabeled data are released (check this link). In addition, the full taxonomy of all the species are also released (check this json file).

Important Dates

Event	Date
Data Released	March 8, 2021
Submission Server Open	March 8, 2021
Submission Deadline	May 31, 2021
Winners Announced	June 7, 2021

Content

• Dataset Intro
• Guidelines
• Data and Annotations
• Submission Format
• Baseline
• Terms of Use
• Acknowledgements

The Semi-Supervised iNaturalist Dataset

This challenge focused on semi-supervised fine-grained classification where we provide labeled data of the target classes and unlabeled data from both target and non-target classes. The data is obtained from iNaturalist, a community driven project aimed at collecting observations of biodiversity. See information below about how this challenge relates to previous iNaturalist challenges at FGVC and other existing datasets.

Our dataset comes with standard training, validation and test sets. The training set consists of:

• labeled images from 810 species, where around 10% of the images are labeled.
• unlabeled images contains unlabeled images from the same set of classes as the labeled images (in-class), plus the images from a different set of classes as the labeled set (out-of-class). The species are guaranteed to have species at the same phylum level in the labels set. This reflects a common scenario where a coarser taxonomic label of an image can be easily obtained. We also provide the kingdom and phylum annotations for each unlabeled image.

The validation and test set contain 5 and 20 images respectively for each of the 810 categories in the labeled set. The distributions of these images are shown in the table below.

Split	Details	#Classes	#Images
Train	Labeled	810	9,721
Train	Unlabeled	-	313,248
Val	Labeled	810	4,050
Test	Public	810	8,100
Test	Private	810	8,100

The number of images per class follows a heavy-tailed distribution as shown in the Figure above. The validation and test sets are uniformly distributed over the classes.

Evaluation Criteria

The performance will be measured as the per-image accuracy (top-1) on the test set. The images in the test set are guaranteed to be one of the 810 categories in the training set.

Differences from Semi-Supervised iNaturalist Challenges at FGVC7

This challenge differs from the previous Semi-Supervised iNaturalist-Aves Competition at FGVC7 (Semi-Aves) in the following ways:

1. Combined unlabeled data: We combined unlabled data from in-class and out-of-class, and do not provide the label.
2. New species: Our dataset has no overlapping species between previous iNat competitions, including iNat-17, iNat-18, iNat-19, iNat-21, and Semi-Aves.
3. Species from different kingdoms: Unlike Semi-Aves where all the species are from the Aves (birds) kingdom, we have a diverse set of species including animal, plants, and fungi. The distribution is shown in the following table.

Kingdom	#Classes
Animalia	429
Plantae	342
Fungi	39

Differences from Existing Semi-Supervised Benchmarks

Unlike prior benchmarks on semi-supervised learning (other than Semi-Aves), our dataset exposes the following challenges:

1. Long-tailed distribution of classes.
2. The unlabeled data come from both target and non-target classes.
3. Fine-grained similarity between classes.

Challenge Guidelines

Participants are allowed to use the labeled and unlabeled images in the training and validation set in any way they like. Participants are not allowed to use images, labels, or pre-trained models from previous iNaturalist competitions (including iNat-17, iNat-18, iNat-19, iNat-21, and Semi-Aves) or any other datasets. However, participants are allowed to use ImageNet-1k (not 21k) pre-trained model (e.g., pre-trained models from torchvision). If you cannot verify the training set of a model, we strongly urge you to not use them because of the potential overlap of class labels.

The general rule is that participants should only use the provided images and annotations (except for the ImageNet-1k pre-trained models) to train a model. For example, dataset from previous iNaturalist competitions or other existing datasets, collecting data from the web or iNaturalist website, or additional annotation on the provided images is not permitted.

Data and Annotations

All the images are stored in JPEG format and have a maximum dimension of 300px. Note this is different from the previous year where the images have a max dimension of 500px.

You can download the dataset files here or on Kaggle:

• Labeled training images [164MB]

• Labeled validation images [67MB]

• Unlabeled training images [5.2GB]

• Test images [271MB]

• Annotations (v2) [31MB]

The annotations are stored in the JSON format and are organized as follows:

{
  "info" : info,
  "annotations" : [annotation],
  "unlabeled_taxa" : [unlabeled_taxa]
}

info{
  "year" : int,
  "version" : str,
  "description" : str,
  "contributor" : str,
  "url" : str,
  "date_created" : datetime,
  "version" : str
}

annotation{
  "class_id" : int,
  "kingdom" : str,
  "phylum" : str,
  "class" : str,
  "order" : str,
  "family" : str,
  "genus" : str,
  "species" : str
}

unlabeled_taxa{
  "image_name" : str,
  "kingdom" : str,
  "phylum" : str
}

For labeled training images in l_train and validation images in val, the subfolder name corresponds to the class_id in the annotation.

For each unlabeled training image, we provide its kingdom and phylum information.

Submission Format

The submission format for the Kaggle competition is a csv file with the following format:

Id,Category
0001feff099c2af9414beb139ff73cbe.jpg,25 198 492 519 412
00027a7dcfe4787085885c3bd0d92da2.jpg,242 213 301 333 802
0008dce7c1c19d6ac8d60246e73c064a.jpg,557 83 152 441 39

The Id corresponds to the file name of the test image, and the Category corresponds to the top five predicted categories separated by space. You should have one row for each test image. Although we only use top-1 accuracy for the leaderboard, the top-5 predictions will be used for more detailed analysis by the competition organizers. A sample solution is provided.

Baseline Performace

The performance using a fine-tuned ImageNet-1k pre-trained ResNet-50 on 224x224 images on the labeled data only is shown below. We hope you can significantly improve the performance using the unlabeled data.

Training Data	Method	Top-1 Val. Accuracy (%)
Labeled train images	ResNet-50 fine-tune	31.0

Terms of Use

By downloading this dataset you agree to the following terms:

1. You will abide by the iNaturalist Terms of Service
2. You will use the data only for non-commercial research and educational purposes.
3. You will NOT distribute the above images.
4. UMass Amherst makes no representations or warranties regarding the data, including but not limited to warranties of non-infringement or fitness for a particular purpose.
5. You accept full responsibility for your use of the data and shall defend and indemnify UMass Amherst, including its employees, officers, and agents, against any and all claims arising from your use of the data, including but not limited to your use of any copies of copyrighted images that you may create from the data.

Citation

If you are using this dataset, please cite this arXiv paper:

@misc{su2021semi_iNat,
      title={The Semi-Supervised iNaturalist Challenge at the FGVC8 Workshop}, 
      author={Jong-Chyi Su and Subhransu Maji},
      year={2021},
      eprint={2106.01364},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgements

FGVC team, Kaggle team, iNaturalist, Grant van Horn, and Oisin Mac Aodha.

Organizer

Jong-Chyi Su (UMass Amherst) jcsu@cs.umass.edu

Previous Competition

• 2020 Competition