Decoupled Pipeline with Reranking and Score Fusion for PU Marine Detection

In the field of marine species detection and classification, positive-unlabeled (PU) datasets present a complex challenge: training samples are sparse and the hidden test set follows a different distribution than the training set. This scenario demands modular architectures that separate proposal generation from fine-grained classification, allowing each stage to be optimized independently. The decoupled pipeline with reranking and fusion, inspired by systems like those developed for FathomNetCLEF 2026, demonstrates how to combine a frozen generic detector (YOLOv8x) acting as a class-agnostic proposal generator, hybrid global and tiled inference with edge filtering, and a vision transformer classifier (DINOv3) fine-tuned with LoRA, all topped by weighted geometric fusion of confidences. A lightweight locally-trained validity head provides an additional reranking signal, improving public validation metrics even if not always the private ones. The most valuable lesson from these experiments is that training-derived and detector-only metrics are not reliable for model selection; it is necessary to use proxy datasets, leaderboard feedback, and targeted ablations. This decoupled approach preserves proposal recall, avoids aggressive filtering, and improves downstream ranking, proving more effective than retraining the detector or using noisy pseudo-labels. From a business perspective, Q2BSTUDIO applies similar principles in the development of custom software for computer vision, where separating detection and classification responsibilities allows scaling systems on the cloud (AWS/Azure) and maintaining cybersecurity for sensitive data. The ability to run tiled inference in parallel using cloud services reduces computational costs, while integrating BI/Power BI enables real-time monitoring of each pipeline component's accuracy. AI agents orchestrate the stages, adjusting confidence thresholds based on detected distribution drift. In a real marine species classification project, a decoupled pipeline like the one described allows updating the classifier without retouching the detector, saving weeks of retraining. Weighted fusion of confidences, combined with validity-based reranking, increases final accuracy by 5% to 10% in PU-labeled environments. Q2BSTUDIO has implemented similar solutions for clients in the environmental sector, achieving robust systems against data scarcity and domain shifts. The key is not to blindly trust internal validation metrics, but to validate with representative real-world proxy sets. This modular design and external verification philosophy is the same one driving the company's cloud AWS/Azure services, where infrastructure adapts to variable inference load. Additionally, cybersecurity is integrated from design, protecting both endangered species data and trained models. The main lesson from this decoupled pipeline is that under PU labeling and distribution shift, it is better to decouple and rerank than to force joint training. This translates into a competitive advantage for companies adopting modular architectures, like those Q2BSTUDIO develops custom. The combination of frozen detectors, LoRA-fine-tuned classifiers, tiled inference, and geometric fusion is not only applicable to marine biology but also to industrial inspection, environmental surveillance, or quality control in aquaculture. In each case, the pipeline is customized to client requirements, integrating explainable AI modules, Power BI dashboards, and real-time alerts. Q2BSTUDIO offers precisely that adaptability, backed by decades of experience in custom software development, artificial intelligence, and cloud computing. Reranking via lightweight validity heads adds an extra refinement layer that, even if not always improving private leaderboard performance, provides greater stability in production environments. Therefore, the company always recommends performing ablation studies before deploying systems in production. The decoupled pipeline analyzed here serves as a conceptual reference for any detection and classification problem with incomplete data and domain shifts. Its practical implementation requires a multidisciplinary team skilled in computer vision, MLOps, and cloud — exactly the profile that Q2BSTUDIO brings together in its custom AI projects. In summary, the strategy of separating proposal generation from classification, applying reranking, and fusing confidences in a weighted manner, validated with proxy sets and external feedback, is a proven recipe for robust systems in PU scenarios. And on that path, Q2BSTUDIO stands as the technology partner that turns this theory into operational, scalable, and secure solutions.