The CLEAR Benchmark

The CLEAR Benchmark: Continual LEArning on Real-World Imagery

By Carnegie Mellon University and CMU Argo AI Center
CLEAR is a novel continual/lifelong benchmark that captures real-world distribution shifts in Internet image collection (YFCC100M) from 2004 to 2014.
For long, researchers in continual learning (CL) community have been working with artificial CL benchmarks such as "Permuted-MNIST" and "Split-CIFAR", which do not align with practical applications. In reality, distribution shifts are smooth, such as natural temporal evolution of visual concepts.
Below are examples of classes in CLEAR-100 that changed over the past decade:
Back to 2004, we had bulky desktop, old-fashioned analog watches, and 2D pixel-art game. Nonetheless, visual concepts gradually evolved from 2004 to 2014, e.g., fancier-looking Macbook Pro, digital watches, and 3D realistic-graphics games.

About CLEAR Benchmark

The CLEAR Benchmark and the CLEAR-10 dataset are first introduced in our NeurIPS 2021 paper.
In spirit of the famous CIFAR-10/CIFAR-100 benchmarks for static image classification tasks, we also collected a more challenging CLEAR-100 with a diverse set of 100 classes.
We hope our CLEAR-10/-100 benchmarks can be the new "CIFAR" as a test stone for continual/lifelong learning community.
We are also extending CLEAR to an ImageNet-scale benchmark. If you have feedback and insights, feel free to reach out to us!
Please cite our paper if it is useful for your research:
title={The CLEAR Benchmark: Continual LEArning on Real-World Imagery},
author={Lin, Zhiqiu and Shi, Jia and Pathak, Deepak and Ramanan, Deva},
booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},

1st CLEAR challenge on CVPR 2022

In June 2022, the 1st CLEAR Challenge was hosted on CVPR 2022 Open World Vision Workshop, with a total of 15 teams from 21 different countries and regions partcipating. You may find a quick summary of the workshop in the below page:
Given the top teams' promising performance on CLEAR-10/-100 benchmarks via utilizing methods that improve generalization, such as sharpness aware minimization, supervised contrastive loss, strong data augmentation, experience replay, etc., we believe there are still a wealth of problems in CLEAR for the community to explore, such as:
  • Improving Forward Transfer and Next-Domain Accuracy
  • Unsupervised/Online Domain Generalization
  • Self-supervised/Semi-Supervised Continuous Learning
In the following pages, we will explain the motivation of CLEAR benchmark, how it is curated via visio-linguistic approach, its evaluation protocols, and a walk-through of the 1st CLEAR Challenge on CVPR'22.
You can also jump to the links for downloading CLEAR dataset: