Research map / Detection and segmentation / Open-set and few-shot detection
Neptune-X: Active X-to-Maritime Generation for Universal Maritime Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Neptune-X, a layout- and caption-conditioned Stable Diffusion variant with bidirectional object-water attention for maritime image synthesis, plus attribute-correlated active sampling of synthetic data to augment detectors.
Problem
Maritime object detection suffers from scarce annotated data and poor generalization across attributes (category, viewpoint, location, imaging environment), and generic layout-to-image models produce physically implausible object-water interactions.
Approach
A multi-condition (text, object box+mask, water mask) generator with Bidirectional Object-Water Attention that models object-water boundary interactions, combined with an active sampling strategy that ranks synthetic samples by attribute-weighted training difficulty (ATDF) estimated from a pretrained detector.
Key results
- Image generation on MGD: FID 18.05 and CAS 79.34, best among compared methods (GLIGEN CAS 77.06, FID 20.02).
- YOLO Score improves to 17.08 mAP vs 12.74 for GLIGEN on generated images.
- With AAS, 10,000 selected synthetic samples raise YOLOv10 mAP from 39.99 to 43.62 (mAP50 61.13 to 65.50), outperforming 10,000 random samples (43.31 mAP).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Neptune-X instead of the strongest baseline for controlled image generation | MGD image generation · MGD | CAS: 77.06 -> 79.34; +2.28 | Helped |
| Use Neptune-X instead of the strongest baseline for controlled object generation | MGD image generation · MGD | YOLO Score mAP: 12.74 -> 17.08; +4.34 | Helped |
| Add generated data to fine-tuning | Grounding DINO · Grounding DINO test set | mAP: 65.03 -> 68.04 (+4.63%) | Helped |
| Add generated data to fine-tuning | Grounding DINO · Grounding DINO test set | mAP50: 86.12 -> 89.86 (+4.34%) | Helped |
| Add water cross-attention to ObjCA (ablation) | Neptune-X generator · MGD image generation | FID: 21.44 -> 19.57 | Helped |
| Add Wat2ObjCA to ObjCA + WatCA (ablation) | Neptune-X generator · MGD image generation | YOLO Score mAP: 13.37 -> 15.60 | Helped |
| Add Obj2WatCA to ObjCA + WatCA (ablation) | Neptune-X generator · MGD image generation | YOLO Score mAP: 13.37 -> 12.52 | Hurt |
| Add both bidirectional attention paths to ObjCA + WatCA (ablation) | Neptune-X generator · MGD image generation | YOLO Score mAP: 13.37 -> 17.08 | Helped |
| Add bidirectional attention paths to ObjCA + WatCA (ablation) | YOLOv10 with Neptune-X generated data · YOLOv10 detection | mAP: 41.09 -> 43.62 | Helped |
| Use AAS instead of random sampling at 5,000 samples (ablation) | YOLOv10 · YOLOv10 detection | mAP: 41.48 -> 43.11 | Helped |
| Use AAS with 5,000 samples instead of random sampling with 10,000 (ablation) | YOLOv10 · YOLOv10 detection | mAP: 43.31 vs 43.11 | No clear effect |
| Add location and imaging-environment ATDF dimensions to viewpoint ATDF (ablation) | AAS · YOLOv10 detection | mAP: 40.26 -> 41.40 | Helped |
Surprising findings
- Detection accuracy under sunny conditions was lower, attributed to water surface glare that is uncommon in training data.
Limitations
- Relies on a fixed, predefined set of discrete attribute categories to estimate training difficulty, limiting granularity.
- Performance saturates when increasing synthetic samples from 10,000 to 20,000.
Experimental setup
Datasets: MaSTr1325 USVInland MIT Sea Grant SMD Seaships Seagull Fvessel LaRS MGD
Benchmarks: MGD test split
Metrics: FID CAS YOLO Score (mAP/mAP50/mAP75) mAP mAP50
Baselines: SD1.5 LayoutDiff GLIGEN InstDiff RC-L2I Random sampling Entropy Variance Margin Greedy K-Center K-Means Corset
Builds on: Stable Diffusion GLIGEN AeroGen RC-L2I InstDiff
Closest papers by meaning
- A Unified Image-Dense Annotation Generation Model for Underwater Scenes (CVPR 2025 · similarity 0.83)
- NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding (NeurIPS 2025 · similarity 0.83)
- Generative Region-Language Pretraining for Open-Ended Object Detection (CVPR 2024 · similarity 0.82)
- Diving into Underwater: Segment Anything Model Guided Underwater Salient Instance Segmentation and A Large-scale Dataset (ICML 2024 · similarity 0.82)
- LAqua: Laplacian Pyramids for Aquatic Segmentation (Student Abstract) (AAAI 2025 · similarity 0.81)
- ODGEN: Domain-specific Object Detection Data Generation with Diffusion Models (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.