README.md

metadata

license: mit

Continually Adapt or Not (CAN) Benchmark

The CAN Benchmark is a curated ICICLE benchmark designed to evaluate the performance of pre-trained models and support the development of adaptation algorithms in the camera trap domain. By providing a structured, temporally-split dataset, CAN enables research on continual adaptation, domain shifts, and long-term model robustness.

Dataset Structure

The dataset consists of two primary components:

1. images/:
   Contains all raw images from the camera trap dataset (CDB-D06).

2. 30/:
   Contains JSON files that divide the dataset into 30-day intervals to support continual learning evaluation:

   • train.json: Training data split by 30-day intervals
   • train-all.json: All training data combined
   • test.json: Test data split by 30-day intervals

This setup allows researchers to simulate real-world temporal data streams in camera trap applications.

How to Use

Clone or download the dataset using:

git lfs install
git clone https://huggingface.co/datasets/ICICLE-AI/CAN_Benchmark

# Unzip the provided archive
unzip CAN_Benchmark/CDB_D06.zip -d CAN_Benchmark/data

You will get the following structure:

CAN_Benchmark/
├── data/
│   ├── images/
│   └── 30/
│       ├── train.json
│       ├── train-all.json
│       └── test.json