🥭 FruitBench: A Multimodal Benchmark for Fruit Growth Understanding

Paper: FruitBench: A Multimodal Benchmark for Comprehensive Fruit Growth Understanding in Real-World Agriculture
Conference: NeurIPS 2025 (submitted)
Authors: Jihao Li*, Jincheng Hu*, Pengyu Fu*, Ming Liu, et al.

---

📌 Dataset Summary

FruitBench is the first large-scale multimodal benchmark designed to evaluate vision-language models on real-world agricultural understanding. It focuses on fruit growth modeling, supporting:

• 🍎 Fruit Type Classification
• 🌱 Growth Stage Recognition (unripe, pest-damaged, mature, rotten)
• 🌾 Agricultural Action Recommendation (keep for further growth, picking it, try to recover it, discard it)
• 🍽️ Consumer Score Prediction (1–100)

The dataset contains 3,200 high-quality expert-annotated images covering 16 fruit categories, each across 4 growth stages.

---

🔍 Tasks

1. Type Classification
2. Growth Stage Identification
3. Action Recommendation
4. Consumer Score Prediction

All tasks are evaluated under both zero-shot and one-shot settings using multimodal large language models (MLLMs).

---

📂 Data Structure

The dataset is organized as follow:

FruitBench/
├── Data/
│ ├── Apple/
│ │ ├── Mature/
      ├──0001.png
      ├──0002.png
      ├──0003.png
      ├──...
      └──0050.png
│ │ ├── Unripe/
│ │ ├── Rotten/
│ │ └── Pest-damage/
│ ├── Banana/
│ ├── Mango/
│ └── ...
├── label/
│ ├── Apple_dataset.parquet
│ ├── Banana_dataset.parquet
│ ├── Mango_dataset.parquet
│ └── ...
├── json/
│ ├── Apple.json
│ ├── Banana.json
│ └── ...

Evaluation

We evaluate a total of 15 multimodal models of different types and sizes, covering diverse model architectures, parameter scales, and vision-language capabilities. The evaluated models include:

• CogVLM2-Llama3-Chat
• DeepSeek-VL-Chat
• DeepSeek-VL2
• InternVL2_5
• Janus-Pro
• Mantis-siglip-llama3
• Mantis-Idefics2
• MiniCPM-Llama3-V2_5
• MiniCPM-o-2.6
• mPLUG-OWL3
• Qwen2.5-VL-Instruct
• Yi-VL
  (15 models in total, with various types and sizes)

⚙️ Environment Setup

We provide both conda and pip setup options (Python 3.11 recommended).

✅ Option A: Conda (Recommended)

conda env create -f environment.yml
conda activate fruitbench

✅ Option B: pip

pip install -r requirements.txt

---

🚀 Usage

1. Clone the Repository

git lfs install
git clone https://huggingface.co/datasets/TJIET/FruitBench

3. Evaluate Models

As an example, the evaluation command for CogVLM2-Llama3-Chat is:

python scripts/CogVLM2-0-shot.py

📊 Benchmark Details

• ✅ 3,200 annotated fruit images
• 📦 16 fruit types: strawberry, tomato, guava, dragon fruit, orange, pear, lychee, mango, kiwi, papaya, apple, grape, pomegranate, peach, banana, pomelo
• 🌱 4 growth stages: unripe, pest-damaged, mature, rotten
• 🧑‍🌾 Expert action labels: keep for growth / pick it / recover / discard
• 🎯 Consumer scores: average of 30 human ratings (range: 1–100)

---