AI Paper Atlas

Research map / Detection and segmentation / Open-set and few-shot detection

Nereus: A Large-Scale Underwater Dataset for Fine-Grained Attribute Understanding and Grounded Counting Perception

Jingtao Deng, Zihao Jin, Yuehe Chen, Feixiang He, Shuiwang Li, Dan Zeng

NeurIPS 2026 · Detection and segmentation › Open-set and few-shot detection · approach: Remote sensing segmentation and change · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 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.

Introduces Nereus, a large-scale underwater dataset aimed at fine-grained attribute understanding and grounded counting perception, motivated by the coarse category labels in existing underwater datasets.

Problem

Existing underwater imagery datasets provide coarse category labels, which do not support recognizing fine-grained biological traits or reasoning over spatially grounded evidence needed for science communication, fishery monitoring, ecological conservation, and oceanographic research.

Approach

A large-scale underwater dataset that annotates fine-grained biological attributes and supports grounded counting, going beyond coarse category labels. Full details unavailable beyond the abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
fine-tune on Nereus instead of NautDataLLaVA-OneVision-7B · Nereus attribute testoverall accuracy (%): 11.27 -> 51.44Improved
fine-tune on Nereus instead of NautDataInternVL2.5-8B · Nereus attribute testoverall accuracy (%): 15.13 -> 51.88Improved
fine-tune on Nereus instead of NautDataQwen2.5-VL-7B · Nereus attribute testoverall accuracy (%): 17.53 -> 46.17Improved
fine-tune on MarineInst20M instead of zero-shotInternVL2.5-8B · Nereus attribute testoverall accuracy (%): 13.41 -> 11.67Hurt
use structured attribute set supervision instead of single-attribute QA pairs (ablation)Qwen2.5-VL-7B · Nereus attribute testoverall accuracy (%): 39.08 -> 41.55Improved
use structured attribute set supervision instead of object-level captions (ablation)Qwen2.5-VL-7B · Nereus attribute testoverall accuracy (%): 16.78 -> 41.55Improved
add degradation type and level as textual priors (ablation)Qwen2.5-VL-7B structured SFT · Nereus attribute testoverall accuracy / semantic similarity: Acc 41.55 -> 40.51; Sim 73.10 -> 74.60Mixed
add DTIUIE image enhancement on top of degradation prior (ablation)Qwen2.5-VL-7B structured SFT · Nereus attribute testoverall accuracy: 41.55 -> 37.36Hurt
add count-only SFT (ablation)InternVL2.5-8B · Nereus counting Track AMAE: 87.38 -> 31.72Improved
add region, grid and visibility QA supervision to count-only SFT (QAS) (ablation)InternVL2.5-8B · Nereus counting Track AMAE: 31.72 -> 25.99Improved
add region/grid QA supervision to count-only SFT (ablation)InternVL2.5-8B · Nereus counting Track Adensest cell accuracy (%): 20.81 -> 52.70Improved
add Track-B density supervision to QAS (Full) (ablation)InternVL2.5-8B · Nereus counting Track AMAE: 25.99 -> 30.48Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy