FishDet-M: A Unified Large-Scale Benchmark for Robust Fish Detection

Overview

FishDet-M is the largest unified benchmark dataset for fish detection in diverse underwater environments, designed to advance ecological monitoring, aquaculture automation, and marine robotics.

Key Features

• 📊 Dataset Scale: 105,556 images with 296,885 annotated fish instances
• 🌊 Diverse Environments: Marine, brackish, aquarium, and occluded scenes
• 🎯 Unified Format: COCO-style annotations with bounding boxes
• 🤖 28 Models Evaluated: Including YOLOv8-v12, DETR, R-CNN variants, and more
• 🧠 CLIP-Based Selection: Automatic model selection using vision-language alignment

Dataset Composition

FishDet-M integrates 13 public datasets spanning:

• Coral reefs and open water
• Aquaculture tanks and ponds
• Clear to highly turbid conditions
• Single fish to dense schools (1-256 instances per image)
• Image resolutions: 78×53 to 4608×3456 pixels

Performance Highlights

Top Performing Models:

• YOLO12x: 0.491 mAP (best overall accuracy)
• YOLO11l: 0.484 mAP with excellent efficiency
• YOLOv8n: 251 FPS for real-time applications

CLIP-Guided Selector:

• Dynamic model selection: 0.444 mAP at 80 FPS
• Zero-shot adaptation to diverse scenes

Key Contributions

1. Unified Benchmark: First large-scale dataset harmonizing diverse underwater fish detection datasets
2. Comprehensive Evaluation: Systematic comparison of 28 state-of-the-art detection models
3. Adaptive Selection: CLIP-based framework for context-aware model deployment
4. Deployment Guidelines: Performance insights across occlusion, scale, and visibility conditions

Applications

• 🐠 Ecological monitoring and biodiversity assessment
• 🎣 Aquaculture automation and fish counting
• 🤖 Underwater robotics and autonomous vehicles
• 🔬 Marine conservation research

Citation

@article{abujabal2025fishdetm,
  title={FishDet-M: A Unified Large-Scale Benchmark for Robust Fish Detection and CLIP-Guided Model Selection in Diverse Aquatic Visual Domains},
  author={Abujabal, Muayad and Saad Saoud, Lyes and Hussain, Irfan},
  journal={IEEE Transactions on Image Processing},
  year={2025}
}

Resources

• Project Website: https://lyessaadsaoud.github.io/FishDet-M/
• Dataset & Code: Available through the project page
• Pretrained Models: All 28 evaluated models with checkpoints
• Interactive GUI: Desktop application for model comparison

Acknowledgments

This work was supported by the Khalifa University Center for Autonomous Robotic Systems (KUCARS) under Awards RC1-2018-KUCARS and CIRA Awards 8474000419 and 8434000534.

---

Khalifa University, Abu Dhabi, UAE