Voxel51/WebUOT-238-Test

Dataset Card for WebUOT-238-Test

This is a FiftyOne dataset with 238 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/WebUOT-238-Test")

# Launch the App
session = fo.launch_app(dataset)

Dataset Description

WebUOT-1M is the largest million-scale benchmark for underwater object tracking (UOT), designed to address limitations in existing datasets by providing diverse underwater scenarios, rich annotations, and language prompts. It comprises 1.1 million frames across 1,500 underwater videos, covering 408 target categories categorized into 12 superclasses (e.g., fish, molluscs, inanimate objects). The dataset includes high-quality bounding box annotations, 23 tracking attributes (e.g., illumination variation, camouflage), and language descriptions for multimodal tracking research.

Note: This dataset, which has been parsed into FiftyOne format, comprises 238 randomly selected videos from the WebUOT-1M test set for a total of 192,000+ frames.

Dataset Details

• —Curated by: Chunhui Zhang (Shanghai Jiao Tong University), Li Liu (HKUST-Guangzhou), Guanjie Huang (HKUST-Guangzhou), Hao Wen (CloudWalk), Xi Zhou (CloudWalk), Yanfeng Wang (Shanghai Jiao Tong University).
• —Funded by: National Natural Science Foundation of China (No. 62101351), Key R&D Program of Chongqing (cstc2021jscx-gksbX0032).
• —Language(s): English (annotations and language prompts).
• —License: Creative Commons (intended for academic research).
• —Shared by: Harpreet Sahota, Hacker-in-Residence @ Voxel51

Dataset Sources

• —Repository: https://github.com/983632847/Awesome-Multimodal-Object-Tracking/tree/main/WebUOT-1M
• —Paper: https://arxiv.org/abs/2405.19818

Uses

Direct Use

• —Training/evaluating UOT algorithms.
• —Multimodal tracking (vision + language prompts).
• —Studying domain adaptation (underwater vs. open-air environments).
• —Marine conservation, underwater robotics, and search/rescue applications.

Out-of-Scope Use

• —Non-underwater tracking tasks (e.g., aerial/terrestrial tracking).
• —Commercial applications without proper licensing.
• —Non-visual tasks (e.g., audio analysis).

Dataset Structure

• —Fields:
• —Videos: 1,500 clips (1,020 train / 480 test).
• —Annotations: Bounding boxes, absent labels, 23 attributes (e.g., low visibility, similar distractors).
• —Language Prompts: Text descriptions of targets (e.g., "red clownfish in yellow coral").
• —Metadata: Object categories (408), superclasses (12), resolution, duration.
• —Splits: Train/Test sets divided by videos, ensuring no overlap in categories or scenarios.

Dataset Creation

Curation Rationale

To bridge the gap in UOT research caused by small-scale datasets, WebUOT-1M was created to enable robust model training/evaluation, domain adaptation, and multimodal tracking in complex underwater environments.

Source Data

Data Collection and Processing

• —Sources: YouTube, Bilibili (filtered for diversity).
• —Processing:
• —Manual selection of moving targets.
• —Semi-supervised enhancement for blurry/low-visibility frames.
• —Professional annotation team for bounding boxes and attributes.
• —Final verification by authors.

Who are the source data producers?

Videos were captured by divers, underwater robots, and hobbyists using varied devices (cameras, phones).

Annotations

Annotation Process

• —Tools: In-house annotation tools; enhanced frames for challenging cases.
• —Guidelines: Focus on target motion, bounding box accuracy, and attribute labeling (23 attributes).
• —Validation: Multiple rounds of correction by authors.

Who are the annotators?

A professional labeling team and the authors performed verification.

Citation

BibTeX:

@article{zhang2024webuot,
  title={WebUOT-1M: Advancing Deep Underwater Object Tracking with A Million-Scale Benchmark},
  author={Zhang, Chunhui and Liu, Li and Huang, Guanjie and Wen, Hao and Zhou, Xi and Wang, Yanfeng},
  journal={arXiv preprint arXiv:2405.19818},
  year={2024}
}

Glossary

The following glossary details the attributes of each video.

Here's the content parsed as a markdown table: