YUV20K: A Complexity-Driven Benchmark and Trajectory-Aware Alignment Model for Video Camouflaged Object Detection

Yiyu Liu; Shuo Ye; Chao Hao; Zitong Yu

cs.CVApr 11, 2026

YUV20K: A Complexity-Driven Benchmark and Trajectory-Aware Alignment Model for Video Camouflaged Object Detection

Authors: Yiyu Liu, Shuo Ye, Chao Hao, Zitong Yu

Organizations: Yiyu Liu, Shuo Ye, Chao Hao and Zitong Yu are with the Department of Computer Science, Great Bay University, Dongguan, Guangdong, 523000, China

Abstract

Video Camouflaged Object Detection (VCOD) is currently constrained by the scarcity of challenging benchmarks and the limited robustness of models against erratic motion dynamics. Existing methods often struggle with Motion-Induced Appearance Instability and Temporal Feature Misalignment caused by complex motion scenarios. To address the data bottleneck, we present YUV20K, a pixel-level annoated complexity-driven VCOD benchmark. Comprising 24,295 annotated frames across 91 scenes and 47 kinds of species, it specifically targets challenging scenarios like large-displacement motion, camera motion and other 4 types scenarios. On the methodological front, we propose a novel framework featuring two key modules: Motion Feature Stabilization (MFS) and Trajectory-Aware Alignment (TAA). The MFS module utilizes frame-agnostic Semantic Basis Primitives to stablize features, while the TAA module leverages trajectory-guided deformable sampling to ensure precise temporal alignment. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art competitors on existing datasets and establishes a new baseline on the challenging YUV20K. Notably, our framework exhibits superior cross-domain generalization and robustness when confronting complex spatiotemporal scenarios. Our code and dataset will be available at https://github.com/K1NSA/YUV20K

Figures & tables

Figure 1: Top: Misalignment of rigid sampling convolutions. When handling non-rigid moving objects, standard sampling convolutions with fixed sampling grids fail to adapt to the target’s continuous deformation. This mismatch causes the network to inevitably sample irrelevant background regions, indicated by the blue points, rather than the target features highlighted in yellow . This leads to severe background noise accumulation over time. Bottom: Alignment via deformable sampling. By dynamically adjusting the sampling grid, our deformable strategy faithfully tracks the target’s features across frames, ensuring all sampled points remain yellow . This actively avoids background interference, mitigating noise accumulation and ensuring accurate spatiotemporal feature alignment.

Dataset	Venue	videos/frames	species	Class	B.Box	Mask
CAD [ 12 ]	ECCV’16	9/839	6	✓	$\times$	✓
MOCA [ 5 ]	ACCV’20	141/37,250	67	$\times$	✓	$\times$
MOCA-Mask [ 13 ]	CVPR’22	87/22,939	44	✓	$\times$	✓
YUV20K	-	91/24,295	47	✓	✓	✓

Table 1: Comparison of YUV20K with existing VCOD datasets. YUV20K provides the most comprehensive annotations.

Figure 2: Spatial distribution density maps. Compared to CAD and MoCA-Mask, where targets tend to appear in the central area, our YUV20K provides a broader and more uniform spatial distribution. This diverse placement introduces greater challenges for robust segmentation.

Panel of Figure 3 — Figure 3: Overview of the proposed YUV20K dataset. (a) Representative frames showcasing diverse camouflage strategies across aquatic, aerial, and terrestrial environments. (b) A hierarchical sunburst chart illustrating the rich taxonomic diversity of the annotated animal species. (c) Attribute-based comparison with existing datasets (CAD and MoCA-Mask) across six complex scenarios: Large Displacement Motion ( Ldm ), Camera Motion ( CM ), Occlusion ( Occ ), Multiple Objects ( M-Obj ), Hunting ( Hunt ), and Tiny Object ( T-Obj ). YUV20K provides a significantly larger number of challenging videos across all attributes, notably establishing a new benchmark in extreme cases like M-Obj and Hunt, where previous datasets are severely lacking.

Figure 4: Overall architecture of the proposed framework. Given a video sequence, we construct a multi-scale input pyramid ( $0.5\times,1.0\times,1.5\times$ ) and extract hierarchical features via a shared backbone. Then features are refined by self-attention with Motion Feature Stabilization module to capture global context. The next step decouples the representations into two pathways. In the temporal pathway, the 3D Central DIfference Spatial-Temporal Convolution modules extract robust temporal cues resilient to severe motion dynamics. In spatial way, the output of self-attention directly push into Trajectory-Aware Alignment module(TAA). Guided by these temporal features, the TAA module dynamically samples and aligns the spatial features. Finally, a decoder aggregates these aligned representations for accurate camouflage mask prediction.

Figure 5: Detailed architecture of the Trajectory-Aware Alignment (TAA) module. The motion-enhanced temporal feature

\mathbf{F}_{t}

is utilized to predict trajectory offsets

\Delta\mathcal{P}

and a modulation mask

\mathcal{M}

. The offsets dynamically deform the standard sampling grid to accurately sample the static spatial feature

\mathbf{F}_{s}

. The aligned spatial representations are then modulated by

\mathcal{M}

and add back into the temporal stream.

Method	CAD [ 12 ]							MoCA-Mask [ 16 ]							YUV20K (Ours)
Method	$S_{\alpha}$ $\uparrow$	$F_{\beta}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$\mathcal{M}$ $\downarrow$	$E_{m}$ $\uparrow$	mIoU $\uparrow$	mDice $\uparrow$	$S_{\alpha}$ $\uparrow$	$F_{\beta}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$\mathcal{M}$ $\downarrow$	$E_{m}$ $\uparrow$	mIoU $\uparrow$	mDice $\uparrow$	$S_{\alpha}$ $\uparrow$	$F_{\beta}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$\mathcal{M}$ $\downarrow$	$E_{m}$ $\uparrow$	mIoU $\uparrow$	mDice $\uparrow$
Image-based Methods
SINet [ 16 ]	.621	.405	.380	.045	.720	.310	.401	.580	.250	.230	.040	.650	.210	.280	-	-	-	-	-	-	-
ZoomNet [ 20 ]	.652	.430	.410	.040	.750	.340	.440	.600	.280	.250	.035	.670	.230	.310	.775	.645	.599	.039	.791	.535	.615
Video-based Methods
SLT-Net-ST [ 13 ]	.719	.517	.479	.032	.832	.410	.509	.631	.311	.288	.029	.701	.258	.344	.842	.740	.712	.031	.888	.652	.738

Table 2: Quantitative comparison of models trained on MoCA-Mask. All models in this table are trained exclusively on the MoCA-Mask training set and evaluated on the test sets of CAD (cross-domain), MoCA-Mask (in-domain), and our proposed YUV20K (cross-domain). Bold and red indicate the best performance.

\uparrow

denotes higher is better, and

\downarrow

denotes lower is better. The symbol “-” indicates the results are not available.

Method	CAD [ 12 ]							MoCA-Mask [ 16 ]							YUV20K (Ours)
Method	$S_{\alpha}$ $\uparrow$	$F_{\beta}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$\mathcal{M}$ $\downarrow$	$E_{m}$ $\uparrow$	mIoU $\uparrow$	mDice $\uparrow$	$S_{\alpha}$ $\uparrow$	$F_{\beta}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$\mathcal{M}$ $\downarrow$	$E_{m}$ $\uparrow$	mIoU $\uparrow$	mDice $\uparrow$	$S_{\alpha}$ $\uparrow$	$F_{\beta}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$\mathcal{M}$ $\downarrow$	$E_{m}$ $\uparrow$	mIoU $\uparrow$	mDice $\uparrow$
SLT-Net-ST [ 13 ]	.720	.492	.445	.039	.806	.398	.499	.798	.642	.597	.029	.848	.559	.659	.848	.684	.648	.025	.908	.607	.697
EMIP [ 23 ]	.725	.533	.483	.030	.839	.405	.518	.879	.789	.775	.012	.925	.722	.789	.857	.727	.702	.020	.924	.643	. .717
ZoomNext [ 25 ]	.775	.569	.565	.018	.815	.503	.577	.827	.719	.685	.018	.853	.624	.707	.856	.704	.685	.017	.895	.635	.690
Ours	.775	.582	.569	.017	.830	.505	.578	.831	.726	.693	.017	.857	.632	.714	.858	.706	.688	.015	.897	.639	.692

Table 3: Quantitative comparison of models trained on our proposed YUV20K. All models in this table are trained exclusively on the YUV20K training set and evaluated on the test sets of CAD (cross-domain), MoCA-Mask (cross-domain), and YUV20K (in-domain). This demonstrates the strong generalization ability acquired from our dataset. Bold and red indicate the best performance.

\uparrow

denotes higher is better, and

\downarrow

denotes lower is better.

Method	Ldm			CM			Occ			M-Obj			Hunt			T-Obj
Method	$S_{\alpha}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$mIoU$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$mIoU$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$mIoU$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$mIoU$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$mIoU$ $\uparrow$	$S_{\alpha}$ $\uparrow$	$F_{\beta}^{\omega}$ $\uparrow$	$mIoU$ $\uparrow$
SLT-Net-ST [ 13 ]	.855	.745	.689	.885	.802	.747	.877	.784	.716	.733	.542	.474	.782	.624	.547	.639	.364	.307
SLT-Net-LT [ 13 ]	.851	.754	.698	.880	.809	.754	.875	.801	.731	.728	.546	.477	.779	.638	.559	.638	.364	.306
ZoomNet [ 20 ]	.799	.644	.583	.826	.695	.636	.791	.648	.573	.678	.428	.358	.726	.525	.440	.637	.281	.235
ZoomNext [ 25 ]	.841	.738	.675	.876	.793	.734	.901	.842	.768	.773	.611	.539	.824	.699	.624	.691	.426	.366
EMIP [ 23 ]	.832	.709	.647	.859	.758	.695	.863	.768	.689	.721	.520	.447	.736	.552	.468	.629	.321	.271

Table 4: Attribute-based performance on the proposed YUV20K dataset. We evaluate the video-based methods across six challenging motion scenarios: Ldm (Large displacement motion), CM (Camera Motion), Occ (Occlusion), M-Obj (Multiple Objects), Hunt (Hunting), and T-Obj (Tiny Object). We report the structure-measure (

S_{\alpha}

\uparrow

), weighted F-measure (

F_{\beta}^{\omega}

\uparrow

), and mean Intersection-over-Union (

mIoU

\uparrow

) to demonstrate the robustness of different methods. Bold and red indicate the best performance.

Figure 6: Qualitative visual comparison of our proposed approach against state-of-the-art methods. Representative sequences are selected from the YUV20K MoCA, and CAD datasets (top to bottom). From left to right: the original input frames, ground truth (GT) masks, predictions of our model, ZoomNeXT, EMIP, and SLT-Net (Short-term and Long-term variants).

Panel of Figure 7 — Figure 7: Qualitative comparison of video camouflaged object detection. Visual sequences comparing our proposed method against state-of-the-art models (e.g., EMIP, SLT-LT) on challenging scenarios from the YUV20K dataset (Part 1). By actively avoiding background interference through our deformable strategy, our method consistently maintains temporal coherence, predicts sharper boundaries, and effectively suppresses background noise compared to other approaches.

Panel of Figure 8 — Figure 8: Qualitative comparison of video camouflaged object detection (Continued). Additional sequence results under extreme challenges such as heavy occlusion and tiny objects. Our model demonstrates superior robustness in faithfully tracking non-rigid targets across frames.

Method	CAD				MoCA-mask				YUV20K (Ours)
Method	$S_{m}\uparrow$	$F_{m}^{\omega}\uparrow$	$E_{m}\uparrow$	$\mathcal{M}\downarrow$	$S_{m}\uparrow$	$F_{m}^{\omega}\uparrow$	$E_{m}\uparrow$	$\mathcal{M}\downarrow$	$S_{m}\uparrow$	$F_{m}^{\omega}\uparrow$	$E_{m}\uparrow$	$\mathcal{M}\downarrow$
Baseline	.759	.562	.861	.020	.699	.421	.696	.008	.823	.691	.851	.036
+ MFS	.757	.543	.852	.021	.715	.438	.722	.008	.807	.467	.863	.032
+ TAA w/ vanilla 3D	.762	.573	.877	.019	.707	.443	.715	.008	.808	.664	.860	.029
+ TAA	.763	.575	.866	.019	.715	.459	.728	.007	.805	.657	.867	.030
Ours	.771	.584	.871	.019	.700	.432	.733	.008	.858	.780	.920	.035

Table 5: Ablation study on three datasets.

\downarrow

indicates lower is better, others are higher is better. The best results are highlighted in bold .

Explore similar work

Apr 9, 2026

CAMotion: A High-Quality Benchmark for Camouflaged Moving Object Detection in the Wild

Discovering camouflaged objects is a challenging task in computer vision due to the high similarity between camouflaged objects and their surroundings. While the problem of camouflaged object detection over sequential video frames has received increasing attention, the scale and diversity of existing video camouflaged object detection (VCOD) datasets are greatly limited, which hinders the deeper analysis and broader evaluation of recent deep learning-based algorithms with data-hungry training strategy. To break this bottleneck, in this paper, we construct CAMotion, a high-quality benchmark covers a wide range of species for camouflaged moving object detection in the wild. CAMotion comprises various sequences with multiple challenging attributes such as uncertain edge, occlusion, motion blur, and shape complexity, etc. The sequence annotation details and statistical distribution are presented from various perspectives, allowing CAMotion to provide in-depth analyses on the camouflaged object's motion characteristics in different challenging scenarios. Additionally, we conduct a comprehensive evaluation of existing SOTA models on CAMotion, and discuss the major challenges in VCOD task. The benchmark is available at https://www.camotion.focuslab.net.cn, we hope that our CAMotion can lead to further advancements in the research community.

Siyuan Yao, Hao Sun, Ruiqi Yu +3

School of CyberScience and Technology, Sun Yat-sen University, Shenzhen Campus, Shenzhen 518107, China. · the College of Computing and Data Science, Nanyang Technological University, 50 Nanyang Avenue, Singapore 639798. · School of Computer Science (National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876, China.

Camouflaged Object Detection Video Object Detection

Ask PDF

Sep 8, 2025

Phantom-Insight: Adaptive Multi-cue Fusion for Video Camouflaged Object Detection with Multimodal LLM

Video camouflaged object detection (VCOD) is challenging due to dynamic environments. Existing methods face two main issues: (1) SAM-based methods struggle to separate camouflaged object edges due to model freezing, and (2) MLLM-based methods suffer from poor object separability as large language models merge foreground and background. To address these issues, we propose a novel VCOD method based on SAM and MLLM, called Phantom-Insight. To enhance the separability of object edge details, we represent video sequences with temporal and spatial clues and perform feature fusion via LLM to increase information density. Next, multiple cues are generated through the dynamic foreground visual token scoring module and the prompt network to adaptively guide and fine-tune the SAM model, enabling it to adapt to subtle textures. To enhance the separability of objects and background, we propose a decoupled foreground-background learning strategy. By generating foreground and background cues separately and performing decoupled training, the visual token can effectively integrate foreground and background information independently, enabling SAM to more accurately segment camouflaged objects in the video. Experiments on the MoCA-Mask dataset show that Phantom-Insight achieves state-of-the-art performance across various metrics. Additionally, its ability to detect unseen camouflaged objects on the CAD2016 dataset highlights its strong generalization ability.

Hua Zhang, Changjiang Luo

Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, University of Chinese Academy of Sciences

Video Object Segmentation Vision Foundation Model Adaptation Camouflaged Object Detection

Ask PDF

Mar 2, 2026

YCDa: YCbCr Decoupled Attention for Real-time Realistic Camouflaged Object Detection

Human vision exhibits remarkable adaptability in perceiving objects under camouflage. When color cues become unreliable, the visual system instinctively shifts its reliance from chrominance (color) to luminance (brightness and texture), enabling more robust perception in visually confusing environments. Drawing inspiration from this biological mechanism, we propose YCDa, an efficient early-stage feature processing strategy that embeds this "chrominance-luminance decoupling and dynamic attention" principle into modern real-time detectors. Specifically, YCDa separates color and luminance information in the input stage and dynamically allocates attention across channels to amplify discriminative cues while suppressing misleading color noise. The strategy is plug-and-play and can be integrated into existing detectors by simply replacing the first downsampling layer. Extensive experiments on multiple baselines demonstrate that YCDa consistently improves performance with negligible overhead as shown in Fig. Notably, YCDa-YOLO12s achieves a 112% improvement in mAP over the baseline on COD10K-D and sets new state-of-the-art results for real-time camouflaged object detection across COD-D datasets.

PeiHuang Zheng, Yunlong Zhao, Zheng Cui +1

Nanjing University of Aeronautics and Astronautics

Camouflaged Object Detection Real-Time Object Detection Channel Attention

Ask PDF