Dataset & Benchmark for Camouflaged Motion Object Detection

CAMotion

A High-Quality Dataset for Camouflaged Motion Object Detection in the Wild

CAMotion is a large-scale benchmark for studying camouflaged moving objects in realistic videos. It is designed to support robust evaluation, reproducible comparison, and future research on motion-aware camouflage understanding in unconstrained real-world scenes.

Paper GitHub arXiv Download Evaluation

Under Review Research Use Only Last updated:

REPLACE

Videos

REPLACE

Frames

REPLACE

Overview

A benchmark website should answer three questions clearly: what the dataset is, why it matters, and how others can use it.

What is CAMotion?

CAMotion provides curated in-the-wild videos containing camouflaged moving objects with dense annotations. It focuses on realistic situations such as low contrast, background blending, clutter, occlusion, motion ambiguity, and changing illumination.

Target Task

Given a video, the goal is to identify and segment the camouflaged moving object(s) frame by frame. The benchmark can support both segmentation-style pipelines and motion-guided video understanding methods.

Why a new benchmark?

Existing video benchmarks often emphasize salient or clearly visible objects, while camouflaged motion remains much less explored. CAMotion is created to encourage algorithms that can reason about subtle appearance cues, motion consistency, and difficult foreground-background ambiguity.

Realistic in-the-wild scenarios
Pixel-level annotations
Motion-aware camouflage challenges
Standard splits and reproducible evaluation

Key Features

Highlights

This section is inspired by benchmark homepages that quickly summarize the dataset’s value proposition. [oai_citation:1‡vision.cs.stonybrook.edu](https://vision.cs.stonybrook.edu/~lasot/?utm_source=chatgpt.com)

In-the-Wild Videos

Natural scenes with complex backgrounds, realistic motion, and challenging camouflage patterns.

Dense Annotations

Frame-level masks for reliable training, fair comparison, and detailed error analysis.

Benchmark Protocol

Standardized train / val / test splits with clear metrics and reproducible evaluation.

Research Ready

Download links, codebase, evaluation scripts, and citation info in one clean entry point.

Dataset Scale

Statistics

Put the most convincing numbers here. This part should let visitors understand the dataset scale in 5 seconds.

REPLACE

Videos

REPLACE

Frames

REPLACE

Official Splits

Split

Videos

Frames

Notes

Train

REPLACE

Standard training split

Val

REPLACE

Validation / ablation

Test

REPLACE

Held-out evaluation split

Challenge Distribution

Low contrast / appearance ambiguity
Fast motion or camera motion
Occlusion and partial visibility
Complex background clutter
Small targets and scale variation
Illumination change / weather variation

You can later replace this block with a pie chart / bar chart image.

Protocol & Results

Benchmark

Strong dataset websites separate metrics, baselines, and evaluation entry points clearly. GOT-10k and MOSE both emphasize benchmark protocol and evaluation access. [oai_citation:2‡got-10k.aitestunion.com](https://got-10k.aitestunion.com/?utm_source=chatgpt.com)

Metrics

mIoU
F-measure / S-measure
MAE
Temporal consistency (optional)

Protocol

Define the official setting here, such as whether methods may use optical flow, pretraining, or external data.

Evaluation

Add the evaluation server or submission instructions if you plan to keep the test labels private.

Evaluation Server

Leaderboard / Main Results

View full leaderboard

Method

Backbone

Setting

mIoU

MAE

Baseline-A

REPLACE

Sup.

REPLACE

Baseline-B

REPLACE

Sup.

REPLACE

Ours

REPLACE

Sup.

REPLACE

建议你把最重要的 4–8 个方法放这里，更多方法可以跳转到 GitHub 或 evaluation page。

Visual Examples

Challenge Cases & Demo

MOSE 的页面里一个很强的点是把复杂场景类型可视化展示出来，这对说服力很有帮助。 [oai_citation:3‡mose.video](https://mose.video/)

Low Contrast

Foreground and background have highly similar appearance.

Occlusion

Target disappears partially behind cluttered structures.

Motion Ambiguity

Background movement and camera motion interfere with perception.

Small Objects

Tiny camouflaged targets with limited discriminative cues.

Video Teaser

用 1 个总 teaser video 最好。也可以额外放 2–4 个 challenge short clips。

Replace with MP4 embed / YouTube / Bilibili iframe

Access

Download

Download area should be simple and explicit: data, annotations, metadata, code.

Dataset

Videos / frames · REPLACE size

Main visual data package.

Annotations

Masks / labels · REPLACE size

Pixel-level annotations and target masks.

Splits & Metadata

TXT / JSON · REPLACE size

Official train/val/test split files and metadata.

Codebase

GitHub repository

Training, inference, and evaluation scripts.

Usage Notice. CAMotion is released for non-commercial research purposes only. Please keep the directory structure unchanged if you use the official loader, and cite the paper in academic use.

Support

FAQ

How do I evaluate on the test split?

Replace this answer with your official evaluation protocol and submission instructions.

Can I use external training data?

State whether external pretraining or extra annotations are allowed for fair comparison.

Where can I report bugs or dataset issues?

Add a contact email or GitHub issue page here.

Reference

Citation

Please cite CAMotion if you find it useful in your research:

@article{camotion,
  title   = {CAMotion: A High-Quality Dataset for Camouflaged Motion Object Detection in the Wild},
  author  = {Siyuan Yao and Hao Sun and Hai Long and Ruiqi Yu and Jiehong Li and Xiwei Jiang and Yanzhao Su and Wenqi Ren and Xiaochun Cao},
  journal = {Under review},
  year    = {2026}
}

Paper • GitHub