SeamCam: Quantifying Seamless Camouflage via Multi-Cue Visual Detectability

What we built, and why

Three numbers worth remembering

SeamCam is a category-aware ideal-observer metric. It cares about whether evidence integration succeeds, not about what made it fail — so it correctly rewards both standard background-matching camouflage and disruptive coloration, while remaining stable across detector/segmenter backbones (77.5–79.7% across four pipelines).

Camouflage as failed localization

Given an image \(I\), ground-truth mask \(M^*\), and species name \(c\), SeamCam produces category-conditioned proposals, segments each, and asks: what is the best subset of those segments at recovering the animal? Whatever that best subset can achieve is the upper bound on detectability — and one minus that is the score.

This formulation is naturally robust to two common detector pathologies: duplicate proposals don't change the max-union IoU, and noisy weak detections are absorbed because the maximization only ever picks subsets that help.

Figure 3. SeamCam framework. Given an image and species name, we generate category-conditioned proposals via Grounding DINO, apply semantic and confidence gating, and obtain segmentation masks from SAM-2. All proposal subsets are evaluated; the maximum-IoU subset defines detectability \(D\). The camouflage score is \(1 - D\).

What 94 people thought

We ran a two-alternative forced-choice study: participants saw image pairs and chose "which is more camouflaged?". With ties dropped, attention-screened catch trials, and pairs with fewer than three valid responses excluded, we had 2,290 valid pairs and 94 participants.

McNemar's paired test on the 2,271-pair shared subset gives \(\chi^2_{cc} = 300.67\), \(p < 10^{-67}\) — SeamCam agrees with humans where CamOT disagrees roughly 3.2× more often than the reverse. Per-species Wilson intervals place SeamCam above chance for all 20 species in the study.

Figure 5. Per-species accuracy comparison. SeamCam (green) outperforms CamOT (brown) across all 20 categories. Largest gaps appear in Fish (Δ = 0.45), Sea Horse (Δ = 0.39), and Cat (Δ = 0.36); the smallest in Cicada and Toad.

SeamCam as a training signal

We use SeamCam to mine hard negatives for Direct Preference Optimization of an SD-V2 inpainting model: real camouflage images are winners; the highest-scoring synthetic candidates (most convincingly camouflaged among generated alternatives) become losers.

Figure 4. SeamCam-based DPO sample selection. InternVL3-14B generates 12 environmental prompt variants per training image; SD-V2 Inpainting synthesizes candidates; the highest-SeamCam candidate is selected as the hard negative against the real image (winner).

Ablation: training signal on SD-V2 base

DPO (CamOT) gets the highest CamOT score — but that's in-distribution reward hacking. The independent metrics (KID, FID, HPS-v2) all favor DPO (SeamCam), confirming the gains are real.

Comparison on CamFG-1.5K

Figure 6. Qualitative comparison. Given only the foreground mask, SeamCam-DPO produces backgrounds that match each animal's natural habitat. TFill loses color fidelity, RePaint-L generates implausible scenes, LDM over-saturates, LAKE-RED places animals in ecologically mismatched environments.

CamFG-1.5K

Existing camouflage datasets contain images that are already partially camouflaged, cropped, or occluded — confounds that systematically bias the evaluation of generation models. CamFG-1.5K eliminates these confounds by construction.

1,521 high-resolution iNaturalist images, each containing a complete, well-segmented animal with minimal occlusion. The dataset spans more than 1,000 species across six taxonomic groups (insects, reptiles, amphibians, fish, birds, mammals) and multiple biome types. It will be publicly released to support reproducible benchmarking.

Figure 2. Quality comparison between CamFG-1.5K and existing datasets. Existing collections often include subjects that are cropped or partially camouflaged, biasing the evaluation of camouflage models. CamFG-1.5K features clearly visible animals with minimal obstructions, enabling unbiased assessment of generation methods.

Cite this work

If SeamCam, CamFG-1.5K, or our DPO recipe is useful in your research, please cite:

@inproceedings{monsefi2026seamcam,
  title     = {SeamCam: Quantifying Seamless Camouflage via Multi-Cue Visual Detectability},
  author    = {Monsefi, Amin Karimi and Meyarian, Abolfazl and Khurana, Mridul
               and Wang, Shuheng and Navard, Pouyan and Zhang, Cheng
               and Karpatne, Anuj and Chao, Wei-Lun and Ramnath, Rajiv},
  
  year      = {2026}
}