emmilly/iwildcam-wilds-v2

WILDS iWildCam v2.0: mirror of the official archive

An unmodified copy of iwildcam_v2.0.tar.gz, the iWildCam dataset as distributed by the WILDS benchmark (wilds package 2.0.0). It is here because the official download host (CodaLab worksheets) has been unreliable since 2025 and currently serves an expired TLS certificate (wilds issue #171). Use the official source when it works.

Camera-trap photos labelled with 182 animal classes, with the WILDS splits by camera location:

split	images
train	129,809
val	14,961
test	42,791
id_val	7,314
id_test	8,154

val and test are out-of-distribution (unseen camera locations); id_val and id_test are in-distribution. The WILDS leaderboard metric is macro-F1 on test.

What we measured on this copy

• Classes: 182, all present in train. The label distribution is long-tailed: 63 classes have fewer than 10 training images.
• Image sizes (all 203,029 images read): the images are pre-resized; the smallest is 560×448 px and the median is 597×448 px. So a 448×448 crop never needs padding or resizing and covers about 75% of a typical image.
• Training cost, measured on an NVIDIA RTX PRO 6000 Blackwell Server Edition with 20 CPU cores, ResNet-50 at 448 px, batch 16, fp32:
  • the standard WILDS loader with RandAugment delivers 582 images/s with 8 workers (586 images/s without augmentation, for evaluation);
  • the GPU trains ~288 images/s and evaluates ~950 images/s, using ~10 GB of GPU memory;
  • one epoch over train plus scoring all four evaluation splits takes about 10 minutes, so a 36-epoch run is ~6 GPU-hours.

We use this copy to replicate ABSGD (Qi et al., arXiv:2012.06951) on the WILDS iWildCam leaderboard.

Use it with the wilds package

# pip install wilds huggingface_hub   (needs ~25 GB of free disk while extracting)
import tarfile
from pathlib import Path
from huggingface_hub import hf_hub_download

archive = hf_hub_download("emmilly/iwildcam-wilds-v2", "iwildcam_v2.0.tar.gz", repo_type="dataset")
root = Path("data")  # any folder; the dataset ends up in data/iwildcam_v2.0/
with tarfile.open(archive) as tar:
    tar.extractall(root / "iwildcam_v2.0", filter="data")
(root / "iwildcam_v2.0" / "RELEASE_v2.0.txt").touch()  # tells wilds the data is already present

from wilds import get_dataset
dataset = get_dataset("iwildcam", root_dir=str(root), download=False)
train = dataset.get_subset("train")

Check your download against the archive's SHA-256: 05c872f46eee44315f79bc90f4098c69ae3799a2b95116db02292ea1b4e370d2 (11,957,415,473 bytes).

Provenance and checks

• Source: https://worksheets.codalab.org/rest/bundles/0x6313da2b204647e79a14b468131fcd64/contents/blob/ (the URL in wilds 2.0.0), downloaded on 2026-10-19 09:37:40.
• The download was made with TLS certificate checking turned off, because the server's certificate had expired. So the transfer itself is not authenticated. These checks were run on the content before uploading:
• split sizes exactly equal the WILDS v2.0 definition (table above)
• metadata.csv has the expected columns; 182 classes; metadata MD5 ab4905ee40ac45ab66ddb176d473a4b4
• all 203,029 image files listed in metadata.csv are present (0 missing)
• 1000 images sampled deterministically decode without error (0 failed)
• the gzip stream's built-in CRC check passed during extraction
• The archive is byte-for-byte what was downloaded; nothing was added, removed or re-encoded.

License and citation

The data is licensed under the Community Data License Agreement – Permissive 1.0 (CDLA-Permissive-1.0), as stated by the dataset authors. This mirror is not affiliated with the dataset authors or the WILDS team. If you use it, cite:

@article{beery2020iwildcam,
title   = {The iWildCam 2020 Competition Dataset},
author  = {Beery, Sara and Cole, Elijah and Gjoka, Arvi},
journal = {arXiv preprint arXiv:2004.10340},
year    = {2020}
}
@inproceedings{koh2021wilds,
title     = {{WILDS}: A Benchmark of in-the-Wild Distribution Shifts},
author    = {Koh, Pang Wei and Sagawa, Shiori and Marklund, Henrik and Xie, Sang Michael and Zhang, Marvin and
           Balsubramani, Akshay and Hu, Weihua and Yasunaga, Michihiro and Phillips, Richard Lanas and Gao, Irena and
           Lee, Tony and David, Etienne and Stavness, Ian and Guo, Wei and Earnshaw, Berton A. and Haque, Imran S. and
           Beery, Sara and Leskovec, Jure and Kundaje, Anshul and Pierson, Emma and Levine, Sergey and Finn, Chelsea and
           Liang, Percy},
booktitle = {International Conference on Machine Learning (ICML)},
year      = {2021}
}