FishDet-M

A Unified Benchmark for Underwater Fish Detection with CLIP-Guided Model Selection

Muayyad Abujabal, Lyes Saad Saoud, Irfan Hussain

Khalifa University Center for Autonomous Robotic Systems (KUCARS), Abu Dhabi, United Arab Emirates

Paper Code FishDet-M App Dataset

FishDet-M App Demo

Interactive walkthrough of the FishDet-M application showcasing detection and dataset visualization capabilities.

Live demonstration of FishDet-M's core functionalities.

Abstract

FishDet-M introduces the first unified benchmark for underwater fish detection, combining 13 real-world datasets into a standardized COCO format. It supports robust model evaluation across varying aquatic environments and incorporates CLIP-based semantic guidance for intelligent model selection. The benchmark includes 28 state-of-the-art models, highlighting the superior performance of the YOLOv12 family, especially YOLO12x. FishDet-M enables reproducible research and paves the way for real-time marine monitoring, aquaculture automation, and underwater ecological analysis.

Comparative Positioning of Fish Detection Datasets

Comparative Positioning of Fish Detection Datasets
Comparative positioning of major fish detection datasets across four key axes: (a) task complexity vs. modality richness, (b) data volume vs. annotation quality, (c) ecological diversity vs. visual challenge, and (d) deployment readiness vs. reproducibility. Bubble size indicates dataset scale (number of annotated fish instances).

CLIP-Guided YOLO Model Selection

CLIP-guided Selection

Detection Performance Across 28 Models

Detection Results Grid
Detection results from 28 models on four challenging underwater images. Each row shows GT followed by YOLOv8–v12, YNAS, DETR variants, R-CNN family, RetinaNet, FCOS, and MobileNetV2-SSD. Scenarios include camouflaged fish, occlusion, and low-contrast schools.

BibTeX

@article{saoud2025fishdetm,
  title={FishDet-M: A Unified Large-Scale Benchmark for Robust Fish Detection and CLIP-Guided Model Selection in Diverse Aquatic Visual Domains},
  author={Abujabal, Muayyad and  Saad Saoud, Lyes and Hussain, Irfan},
  journal={Submitted to ....},
  year={2025},
  note={...}
}