ICICLE-AI/CAN_Benchmark

---
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:

```bash
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
```