Research map / Detection and segmentation / Open-set and few-shot detection
MARIS: Marine Open-Vocabulary Instance Segmentation
Bingyu Li, Feiyu Wang, Da Zhang, Zhiyuan Zhao, Junyu Gao, Xuelong Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GPEM geometric prior module to the baseline (ablation) | ConvNeXt-L OV segmenter · MARIS in-domain, overall classes | mAP: 52.17 -> 54.99 | Improved |
| add SAIM semantic alignment module to the baseline (ablation) | ConvNeXt-L OV segmenter · MARIS in-domain, overall classes | mAP: 52.17 -> 55.27 | Improved |
| combine GPEM and SAIM (ablation) | ConvNeXt-L OV segmenter · MARIS in-domain, OV classes | mAP: 50.99 -> 54.02 | Improved |
| add underwater prompts instead of generic templates (ablation) | SAIM, ConvNeXt-B · MARIS in-domain, overall classes | mAP: 42.91 -> 43.89 | Improved |
| add similarity-based template selection on top of underwater prompts (ablation) | SAIM, ConvNeXt-B · MARIS in-domain, overall classes | mAP: 43.89 -> 44.54 | Improved |
| use adaptive alpha fusion instead of MLP fusion (ablation) | GPEM, ConvNeXt-B · MARIS in-domain, overall classes | mAP: 43.87 -> 44.54 | Improved |
| use plain addition instead of adaptive fusion (ablation) | GPEM, ConvNeXt-B · MARIS in-domain, overall classes | mAP: 44.54 -> 43.52 | Hurt |
| enlarge geometry encoder from ViT-B to ViT-L (ablation) | MARIS with ConvNeXt-L · MARIS in-domain vs cross-domain, overall classes | mAP: in-domain 55.22 -> 56.71; cross-domain 46.18 -> 43.70 | Mixed |
| enlarge geometry encoder from ViT-B to ViT-L (ablation) | MARIS with ConvNeXt-B · MARIS in-domain, overall classes | mAP: 44.54 -> 44.37 | No clear effect |
| shrink geometry encoder from ViT-B to ViT-S (ablation) | MARIS with ConvNeXt-B · MARIS in-domain, overall classes | mAP: 44.54 -> 42.36 | Hurt |
| train on COCO and test underwater instead of in-domain training | MARIS ConvNeXt-B · MARIS overall classes | mAP: 44.37 -> 32.62 | Hurt |
| use MARIS instead of MAFT+ in cross-domain | ConvNeXt-L · MARIS cross-domain, overall classes | mAP: 40.27 -> 46.18 | Improved |
Closest papers by meaning
- Diving into Underwater: Segment Anything Model Guided Underwater Salient Instance Segmentation and A Large-scale Dataset (ICML 2024 · similarity 0.87)
- Leveraging Depth and Language for Open-Vocabulary Domain-Generalized Semantic Segmentation (NeurIPS 2025 · similarity 0.85)
- Fantastic Animals and Where to Find Them: Segment Any Marine Animal with Dual SAM (CVPR 2024 · similarity 0.85)
- NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding (NeurIPS 2025 · similarity 0.85)
- Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models (CVPR 2024 · similarity 0.84)
- A Unified Image-Dense Annotation Generation Model for Underwater Scenes (CVPR 2025 · similarity 0.84)
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.
