---
title: NICDF-AquaDetSeg
canonical_url: "https://www.modelscope.cn/datasets/NICDFcau/NICDF-AquaDetSeg"
md_url: "https://www.modelscope.cn/datasets/NICDFcau/NICDF-AquaDetSeg.md"
repository: NICDFcau/NICDF-AquaDetSeg
chinese_name: "NICDF 水生目标检测与分割数据集（AquaDetSeg）"
last_updated: 2026-08-16
license: CC-BY-NC-4.0
storage_size: "3.9 GB"
downloads: 239071
stars: 1
---

# NICDF-AquaDetSeg

> NICDF-AquaDetSeg - NICDFcau 在 ModelScope 开源的数据集。A train-only aquatic object detection and segmentation dataset released by the National Innovation Center for Digital Fishery (NICDF), with COCO, semantic-mask and source annotation views.

NICDFcau/NICDF-AquaDetSeg 是 ModelScope 魔搭社区上的数据集，存储大小 3.9 GB，采用 CC-BY-NC-4.0 许可。

- **Repository**: NICDFcau/NICDF-AquaDetSeg
- **License**: CC-BY-NC-4.0
- **Storage size**: 3.9 GB
- **Downloads**: 239071
- **Stars**: 1
- **Last updated**: 2026-08-16

Source: https://www.modelscope.cn/datasets/NICDFcau/NICDF-AquaDetSeg

---

# NICDF Aquatic Object Detection and Segmentation Dataset (AquaDetSeg)

Release: `NICDF-v1.0.0`  
Institutional owner: **National Innovation Center for Digital Fishery (NICDF)**  
Public scope: **train only**

## Overview

A train-only aquatic-object detection and segmentation dataset released by NICDF, providing COCO detection/instance annotations, semantic masks where available, and stable NICDF_taxon_id labels.

This repository is the reviewed ModelScope delivery view for `AquaDetSeg`. It contains training data only; validation, public-test, private-test, and evaluator-only material are not included.

## Dataset scale

- Training media records: 12,818
- Unique `asset_id` values: 12,818
- Unique media paths: 12,818
- Formal species identity: `NICDF_taxon_id` for one species or `NICDF_taxon_ids` for mixed-species records

The COCO view contains 12,659 images and 72,769 annotations. The semantic-mask table contains 221 unique masks. Every published semantic mask is a single-channel uint8 PNG with class indices 0 (background) and 1 (foreground). Antialiased grayscale crab masks were thresholded at >=128, and byte-identical duplicate source occurrences were represented once per asset.

## Repository layout

```text
.
|-- README.md
|-- LICENSE
|-- RELEASE.json
|-- data/train.csv
|-- media/sha256/...
|-- annotations/AquaDetSeg/...
|-- metadata/...
|-- checksums.sha256
`-- checksums.modelscope.sha256
```

Each row in `data/train.csv` represents one training media asset. Core fields are:

- `asset_id`
- `image`
- `sha256`
- `file_size_bytes`
- `media_type`
- `modality`
- `width`
- `height`
- `source_dataset`
- `source_partition`
- `session_id`
- `sequence_id`
- `frame_index`
- `global_split`
- `split_basis`
- `annotation_set`
- `NICDF_taxon_id`
- `NICDF_taxon_ids`
- `taxon_count`
- `taxonomy_status`
- `canonical_name_zh`
- `scientific_name`
- `taxonomic_rank`
- `label`
- `label_name`
- `rights_status`
- `rights_id`
- `data_license`
- `institutional_owner`
- `publication_visibility`
- `has_semantic_mask`
- `semantic_masks`
- `semantic_mask_encoding`
- `semantic_masks_table`
- `coco_annotations`

`image` is a repository-relative path. Missing contextual fields are left empty and are not inferred. Site, batch, pond/tank, and date are not ModelScope-required fields.

## Loading

```python
from modelscope.msdatasets import MsDataset

dataset = MsDataset.load(
    "NICDFcau/NICDF-AquaDetSeg",
    subset_name="default",
    split="train",
)
```

The COCO, semantic-mask, and MOTChallenge files under `annotations/` are the authoritative task views. The platform table preview does not replace task-format parsing.

## Split and leakage policy

This repository exposes only `default/train`. NICDF validation and test partitions remain in the private evaluation chain and are not uploaded. Complete site, batch, session, event, and sequence groups are kept together.

## Taxonomy

Published records use reviewed `NICDF_taxon_id` values and include canonical Chinese names, scientific names, and taxonomic ranks. Koi cultivars and common carp share the species-level identity `Cyprinus carpio`; cultivar information remains in label views and is not presented as a separate species.

## Ownership and licence

- Institutional owner: National Innovation Center for Digital Fishery (NICDF)
- Data licence: CC BY-NC 4.0
- Redistribution scope: train-only release

Use of this dataset must comply with `LICENSE`.

## Citation

Until a related paper or DOI is issued, cite the institutional dataset record:

> National Innovation Center for Digital Fishery. (2026). NICDF Aquatic Object Detection and Segmentation Dataset (AquaDetSeg) (NICDF-v1.0.0) [Dataset]. ModelScope.

## Known limitations

- Only training data are provided; test labels are not public.
- Semantic masks are available only for the subset identified by `has_semantic_mask`.
- Image quality, viewpoint, density, and occlusion vary across sources.
- The same canonical media asset may be reused across NICDF task datasets through a stable `asset_id` and SHA-256.

## Integrity

`checksums.sha256` records the canonical NICDF source bytes before upload. ModelScope may normalize
line endings in text files such as XML and Markdown. For a repository downloaded from ModelScope,
verify against `checksums.modelscope.sha256`, which records the platform-delivered bytes and covers
all release files except itself and the platform-maintained `dataset_infos.json`. Both manifests are
provided so that source provenance and downloadable-byte reproducibility remain independently auditable.