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

Camouflage Anything: Learning to Hide using Controlled Out-painting and Representation Engineering

CVPR 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Open-set and unknown object detection · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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.

Proposes Camouflage Anything, a ControlNet-style controlled out-painting model with contrastive CLIP representation engineering for generating camouflaged images, a new GMM-based optimal-transport camouflage metric (CamOT), and LoRA fine-tuning of BiRefNet on the generated data.

Problem

Generating realistic camouflaged images that hide arbitrary segmented objects, since existing datasets are small and costly and FID/KID do not capture camouflage quality.

Approach

Condition a ControlNet on a control image whose background is replaced by the mean background colour, and refine the CLIP text representation with a contrast-attention term (source minus null statement) to out-paint camouflaged backgrounds; score camouflage with a second-order Wasserstein distance between foreground and boundary-background GMMs (CamOT).

Key results

• On the LAKE-RED test set, the CO+RE-BG variant reaches overall FID 40.53 and KID 0.0155, versus 64.27 and 0.0355 for LAKE-RED.
• Under CamOT, the CO+RE+BG configuration reaches a mean of 0.8403 on SOD and 0.8300 on SEG, versus 0.7154 and 0.6772 for LAKE-RED.
• LoRA fine-tuned BiRefNet improves S-measure on the generated SOD split from 0.722 to 0.916 and on SEG from 0.646 to 0.885, but slightly degrades COD10K S-measure (0.913 to 0.904).

Measured effects reported in the paper

Surprising findings

• The configuration with the best CamOT and FID/KID (CO+RE+BG) has worse FID/KID than CO+RE-BG, which the authors attribute to generated images being more camouflaged than the reference set.
• LoRA tuning of BiRefNet on generated data slightly degrades performance on the real COD10K benchmark.

Limitations

• Struggles to generate images containing people due to the Stable Diffusion 1.5 base model.
• Sometimes places objects in unrealistic scenes.
• FID/KID penalize higher-camouflage generations, so they are an imperfect quality measure; CamOT is not validated against human judgement.

Experimental setup

Datasets: COD10K CAMO NC4K COCO DUTS LAKE-RED test set

Benchmarks: LAKE-RED test set (COD, SOD, SEG splits) COD10K

Metrics: FID KID CamOT S-measure weighted F-measure mean E-measure MAE

Baselines: LAKE-RED TFill LDM RePaint-L LCGNet DCI AdaIN CI AB

Builds on: ControlNet Stable Diffusion LAKE-RED BiRefNet LoRA

Closest papers by meaning

• Deep Camouflage Images (AAAI 2020 · similarity 0.88)
• Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection (CVPR 2025 · similarity 0.88)
• CamSAM2: Segment Anything Accurately in Camouflaged Videos (NeurIPS 2025 · similarity 0.87)
• Discover, Segment, and Select: A Progressive Mechanism for Zero-shot Camouflaged Object Segmentation (CVPR 2026 · similarity 0.87)
• CamoDiffusion: Camouflaged Object Detection via Conditional Diffusion Models (AAAI 2024 · similarity 0.87)
• LAKE-RED: Camouflaged Images Generation by Latent Background Knowledge Retrieval-Augmented Diffusion (CVPR 2024 · similarity 0.85)