ChatGPT and general-purpose AI count fruits in pictures surprisingly well without programming or training

Abstract

• Evaluated ChatGPT's performance in counting fruits from agricultural images. • Foundation model outperformed YOLOv8 in accuracy with minimal training data. • ChatGPT's accuracy improved with user feedback in counting tasks. • LLMs and foundation models greatly reduce time needed for fruit counting. • General-purpose AI tools offer scalable, accessible solutions for smallholder farmers. General-purpose artificial intelligence (AI) can facilitate agricultural digitalization as many tools do not require coding. Yet, it remains unclear how well the emerging general-purpose AI technologies can perform object counting, which is a fundamental task in agricultural digitalization, in comparison to the current standard practice. We show that ChatGPT (GPT4 V) demonstrated moderate performance in counting coffee cherries from images, while the T-Rex, foundation model for object counting, performed with high accuracy. Testing with a hundred images, we examined that ChatGPT can count cherries, and the performance improves with human feedback (R 2 = 0.36 and 0.46, respectively). The T-Rex foundation model required only a few samples for training but outperformed YOLOv8, the conventional best practice model (R 2 = 0.92 and 0.90, respectively). Obtaining the results with these models was 100x shorter than the conventional best practice. These results bring two surprises for deep learning users in applied domains: a foundation model can drastically save effort and achieve higher accuracy than a conventional approach, and ChatGPT can reveal a relatively good performance especially with guidance by providing some examples and feedback. No requirement for coding skills can impact education, outreach, and real-world implementation of generative AI for supporting farmers.

Ähnliche Arbeiten