EUVIMEDEUVIMED entdecken
Crossref

Benchmarking and Fine-Tuning Monocular Depth Models for Fish Biomass Estimation in Aquaculture

Alexandrou A, Christodoulou F, Komninos D, Ozdeger T, Seferis K

Journal of Aquaculture, Marine Biology & Ecology · 2026 · 01 (01)

DOI: 10.37722/jambe.2026103 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Aquaculture, Marine Biology & Ecology, 2026. Verfasst von Alexandrou A und 4 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate and non-invasive fish biomass estimation is essential for efficient feeding, environmental sustainability, and productivity in modern aquaculture. While monocular vision systems offer a low-cost and scalable alternative to stereo cameras, reliable metric depth estimation from a single RGB image remains a significant challenge, particularly in complex underwater environments. This study investigates the feasibility of using state-of-the-art monocular depth foundation models for fish biomass estimation in real aquaculture cages and evaluates the impact of domain-specific fine-tuning. Four advanced monocular depth models, Depth Pro, UniDepth v2, Depth Anything v2, and the recently released Depth Anything v3, were evaluated in a zero-shot setting using real underwater footage collected from commercial sea cages. The initial evaluation showed strong scale ambiguity and systematic overestimation across all models, highlighting the need for explicit depth calibration when applying monocular approaches in underwater environments. A finetuned version of Depth Anything v2 trained on stereo-aligned underwater data delivered markedly better results. In controlled experiments using a reference object of known length, the average length estimation error was reduced from more than 7 cm to 3.84 cm. Across all test videos, the error was further reduced from 3.76 cm to 2.42 cm, demonstrating the clear benefit of underwater-specific fine-tuning. These findings demonstrate that, although off-the-shelf monocular models are limited by scale ambiguity in underwater conditions, domain-adapted monocular depth estimation can achieve reliable performance for practical biomass estimation. This work highlights the strong potential of fine-tuned monocular systems, as scalable and cost-effective alternatives for real-time aquaculture monitoring and precision feeding applications.

Zitieren

APA

A, A., F, C., D, K., T, O., & K, S. (2026). Benchmarking and Fine-Tuning Monocular Depth Models for Fish Biomass Estimation in Aquaculture. Journal of Aquaculture, Marine Biology & Ecology, 01(01). https://doi.org/10.37722/jambe.2026103

Vancouver

A A, F C, D K, T O, K S. Benchmarking and Fine-Tuning Monocular Depth Models for Fish Biomass Estimation in Aquaculture. Journal of Aquaculture, Marine Biology & Ecology. 2026;01(01). doi:10.37722/jambe.2026103

BibTeX
@article{A2026,
  author = {A, Alexandrou and F, Christodoulou and D, Komninos and T, Ozdeger and K, Seferis},
  title = {Benchmarking and Fine-Tuning Monocular Depth Models for Fish Biomass Estimation in Aquaculture},
  journal = {Journal of Aquaculture, Marine Biology & Ecology},
  year = {2026},
  volume = {01},
  number = {01},
  doi = {10.37722/jambe.2026103}
}
Kostenlos starten →