# Reducing annotation effort in agricultural data: simple and fast unsupervised coreset selection with DINOv2 and K-means

## TL;DR

This paper introduces a new method to reduce the need for annotated data in agriculture by using DINOv2 and K-means to select representative samples for annotation.

## Contribution

A novel coreset selection method combining DINOv2 and K-means for efficient and effective data annotation in agricultural deep learning applications.

## Key findings

- The proposed method outperforms random selection with up to 0.15 improvement in F1 score.
- Using DINOv2 as a feature extractor significantly contributes to the method's effectiveness.
- The approach is validated on two agricultural datasets with consistent performance improvements.

## Abstract

The need for large amounts of annotated data is a major obstacle to adopting deep learning in agricultural applications, where annotation is typically time-consuming and requires expert knowledge. To address this issue, methods have been developed to select data for manual annotation that represents the existing variability in the dataset, thereby avoiding redundant information. Coreset selection methods aim to choose a small subset of data samples that best represents the entire dataset. These methods can therefore be used to select a reduced set of samples for annotation, optimizing the training of a deep learning model for the best possible performance. In this work, we propose a simple yet effective coreset selection method that combines the recent foundation model DINOv2 as a powerful feature selector with the well-known K-Means clustering method. Samples are selected from each calculated cluster to form the final coreset. The proposed method is validated by comparing the performance metrics of a multiclass classification model trained on datasets reduced randomly and using the proposed method. This validation is conducted on two different datasets, and in both cases, the proposed method achieves better results, with improvements of up to 0.15 in the F1 score for significant reductions in the training datasets. Additionally, the importance of using DINOv2 as a feature extractor to achieve these good results is studied.