Decoupled Pipeline with Proposal Reranking and Score Fusion for Positive-Unlabeled Marine Species Detection

2026-07-21 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

DS@GT ARC developed a multi-stage system for the FathomNetCLEF 2026 competition, which challenges participants with positive-unlabeled marine species detection under sparse training labels and an out-of-distribution hidden test set. Their approach, which secured 12th place out of 102 teams, utilized a frozen Megalodon YOLOv8x detector for class-agnostic proposal generation. It combined global and tiled inference with tile-edge filtering, then classified expanded proposal crops using a LoRA-finetuned DINOv3 ViT-H classifier. Predictions were finally ranked via weighted geometric fusion of detector and classifier confidence. A key finding was the unreliability of train-derived validation and detector-only metrics for model selection, leading the team to rely on proxy datasets, leaderboard feedback, and targeted ablations. Experiments showed that preserving proposal recall, avoiding over-aggressive filtering, and improving downstream ranking were more effective than detector fine-tuning or direct training on noisy pseudo-labels.

Key takeaway

For Machine Learning Engineers developing positive-unlabeled object detection systems, relying solely on train-derived validation metrics can be misleading. You should prioritize robust validation using proxy datasets and leaderboard feedback to guide model selection. Focus your efforts on preserving proposal recall, carefully managing filtering thresholds, and enhancing downstream prediction ranking through techniques like weighted confidence fusion, rather than just fine-tuning the base detector.

Key insights

For positive-unlabeled object detection, robust validation and effective proposal reranking are crucial given sparse labels and distribution shifts.

Principles

• Train-derived validation can be unreliable.
• Prioritize proposal recall and ranking.
• Avoid over-aggressive filtering.

Method

The system uses a frozen YOLOv8x for proposals, combines global/tiled inference, classifies with a LoRA-finetuned DINOv3 ViT-H, and ranks predictions via weighted geometric fusion of confidences.

In practice

• Use proxy datasets for validation.
• Implement weighted confidence fusion.
• Explore reranking signals like validity heads.

Topics

• Marine Species Detection
• Positive-Unlabeled Learning
• Object Detection Pipelines
• Proposal Reranking
• Score Fusion
• Out-of-Distribution Detection

Code references

• dsgt-arc/fathomnetclef-2026

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Learned Non-Maximum Suppression for 3D Object Detection — Learned filtering modules, D2D-Rescore and GossipNet3D, enhance 3D object detection by replacing heuristic NMS with detection-to-detection relation leveraging.
• Hierarchical Specialised Ensembles for Classification of Zebrafish Phenotypes Using the Selected Image Recognition Methods — Hierarchical specialized ensembles effectively classify zebrafish phenotypes, with ConvNeXt excelling as a backbone.
• Multi-Scale ViT Inference with Habitat-Fit Priors and kNN Retrieval for Multi-Species Plant Identification — Multi-scale ViT inference combined with habitat-fit priors and kNN retrieval significantly improves multi-species plant identification.
• Robust Image Processing Techniques for Construction Environment Monitoring Using Underwater Robots — Explicitly modeling depth-dependent forward scattering and marine snow significantly enhances underwater image processing for robotic construction monitoring.
• Tricking Open-World Object Recognition Models: Uncertainty in Out-of-Distribution Detection — Open-World object recognition models struggle with uncertainty when prompted for non-existing objects, particularly GPT-4o's overconfidence.
• Rethinking Multi-Branch and Cross-Backbone Fusion for Vehicle Re-Identification in the Foundation-Model Era — Multi-branch and cross-backbone fusion for vehicle Re-ID offers no significant gains over a single, well-tuned foundation model.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.