A curated ICICLE benchmark for evaluating the performance of pre-trained models and fostering the development of adaptation algorithms in the camera trap domain.
Tags: Foundation-AI
MIT License. See LICENSE for details.
Lila bc: Labeled information library of alexandria: Biology and conservation. https://lila.science/. 1, 3
National Science Foundation (NSF) funded AI institute for Intelligent Cyberinfrastructure with Computational Learning in the Environment (ICICLE) (OAC 2112606)
To report any issue(s) with this benchmark, please create an issue in this github or contact: jeon.193@osu.edu
Clone or download the dataset using:
git lfs install
git clone https://huggingface.co/datasets/ICICLE-AI/CAN_Benchmark
Then unzip the data archive:
unzip CAN_Benchmark/CDB_D06.zip -d CAN_Benchmark/data
Please see this link ICICLE Camera Trap Github, where you can find detailed instruction on how to use this dataset
The dataset consists of of two folders:
images: You can find all the images from a camera trap CDB-D06'30': You will find three json files in this folder. The whole camera trap dataset has been divided into 30 days interval, for continual learning.train.json: Json file for training data divided into 30 days interval.train-all.json:Json file containing all training data.test.json: Json file containing all test data divided into 30 days interval.
- Total Size: ~2.1 TB
- Total Images: 3,317,354
- Regions Covered (17): apn, cdb, eno, kar, kga, kru, mad, mtz, pln, rua, nz (excluding nz_bad), orinoquia, idaho, serengeti, wellington, caltech, na
- Number of Cameras: 546
- File Format: JPG
- Image Resolution Range: min 56×58, max 1920×2560
- Average Classes per Dataset: ~11
To request access to the full dataset, please contact the maintainers at: jeon.193@osu.edu