ICME 2026 APS Challenge

Understanding ancient diets is essential for reconstructing subsistence strategies and human–environment interactions. Plant seeds are key archaeobotanical evidence, yet their identification still relies on time-consuming expert analysis. Existing datasets mainly contain well-preserved samples and fail to reflect real archaeological conditions.

To address this gap, the APS Challenge extends the original APS dataset by incorporating severely damaged and carbonized seed images. Participants are invited to develop robust fine-grained classification models that handle high inter-class similarity and large intra-class variation under realistic conditions.

Register for the Challenge: To streamline the participation process, each team only needs to register one Codabench account during the competition. The complete list of team members and their affiliations must be declared when submitting the technical report.

Important note: The email address used to register the Codabench account must belong to one of the team members.

The APS Challenge dataset extends the original APS (Ancient Plant Seeds) dataset with severely damaged and carbonized seed images, providing a more realistic representation of archaeological conditions. The dataset includes various seed categories with fine-grained annotations to facilitate the development of robust classification models.

Figure 1: Comparison of the size and damage of ancient plant seeds and the condition of modern seeds. We divide the seeds into four sizes, where each has three columns of seed images. The first column shows the true seed size under 1.6× magnification, the second column shows seeds with pronounced differences caused by uncontrollable factors, and the third column shows the corresponding modern seeds.

We provide a baseline model, APSNet, while allowing participants to adopt any methods beyond the baseline. Our previous work, which has been publicly released on arXiv, is adopted as the baseline method for this challenge and also serves as the source of the APS dataset. The proposed APSNet demonstrates competitive performance on fine-grained ancient plant seed classification. The source code of the baseline method is publicly available at Github. Participants are encouraged to cite our baseline paper if it is relevant to their research.

@article{xing2025towards,
title={Towards Ancient Plant Seed Classification: A Benchmark Dataset and Baseline Model},
author={Xing, Rui and Cong, Runmin and Wu, Yingying and Wang, Can and Tang, Zhongming and Wang, Fen and Wu, Hao and Kwong, Sam},
journal={arXiv preprint arXiv:2512.18247},
year={2025}
}