A camouflage detector outputs a per-pixel score that says "object" or "background." That score hides a distinction that matters. A pixel can score 0.5 because the network sees nothing either way, as on a blank patch of sand. It can also score 0.5 because it sees strong evidence for both answers at once, as when fur texture matches leaf texture. The first case is ignorance. The second is conflict. Camouflage is, almost by definition, the second case: the object looks like the background. Closed-world detectors throw this difference away. They are trained on images that always contain a hidden animal, and they hallucinate one in nearly every empty scene. Is There Really a Camouflaged Object? reports false-positive rates as high as 97.4% on such images.
We propose to recover the distinction from any trained camouflage segmenter, with no retraining. The tool is belief functions (Dempster–Shafer theory), which assign probability mass to sets of outcomes, including the "don't know" set {object, background}. We split the final linear layer's per-channel contributions into evidence for the object and evidence for the background, and combine them with Dempster's rule. The result, at every pixel, is a conflict mass and an ignorance mass alongside the usual probability. Our central claim is that these two maps separate the three image types in realistic camouflage detection without any negative training images. Camouflaged-object images should show high conflict inside the predicted mask. Salient, non-camouflaged objects should show strong one-sided evidence with low conflict. Pure backgrounds should show high ignorance.
Core Claim
Consider a pixel with feature vector $h \in \mathbb{R}^C_{\ge 0}$, the output of a ReLU. The segmentation head computes the logit $z = \sum_c w_c h_c + b$. Following the evidence reading of linear layers used by EUQ for vision-language models, we define two totals. Object evidence is $E^+ = \sum_c \max(w_c h_c, 0)$, and background evidence is $E^- = \sum_c \max(-w_c h_c, 0)$. Each channel is treated as a simple support function, and Dempster's rule combines them. The two totals then give
$$\kappa = (1-e^{-E^+})(1-e^{-E^-}), \qquad m(\{O,B\}) = \frac{e^{-E^+-E^-}}{1-\kappa},$$
where $\kappa$ is the conflict mass and $m(\{O,B\})$ is the ignorance mass left on the "either" set.
The sigmoid probability depends only on the difference $E^+ - E^-$. Conflict and ignorance depend on the sum and the product of the two totals, which is exactly what the sigmoid discards. Two pixels with logit zero can be total ignorance ($E^+=E^-=0$) or near-total conflict ($E^+=E^-=10$). We hypothesize that this discarded dimension is where "is there really a camouflaged object?" can be answered.
Relation to prior camouflage work
Existing uncertainty-aware camouflage work does not look at this dimension. Three systems produce a single uncertainty score:
- UGTR samples a per-pixel Gaussian 50 times.
- PUENet uses Monte Carlo dropout inside a conditional VAE and distils the variance into a fast student network.
- UMBD fuses a Bayesian branch with an attention branch into one mask that decides where to refine.
A fourth, CANet, trains a second U-Net to predict the detector's error map and explicitly argues that epistemic uncertainty is secondary in camouflage.
All four evaluate only on images guaranteed to contain a target, and none separates conflict from ignorance. OPCNet does tackle the open-world problem. It uses a supervised two-stage existence classifier trained on negative images, and its outputs are deterministic scores. Our proposal differs from OPCNet in two ways: the rejection signal comes from the structure of the existing head rather than from negative supervision, and it has two dimensions rather than one.
Relation to uncertainty theory
The uncertainty literature both supplies the method and constrains it. EUQ supplies the decomposition. EUQ showed, in vision-language models, that conflict tracks hallucination (AUROC 0.761 on hallucination) and ignorance tracks out-of-distribution inputs. It has never been applied to dense prediction or to a domain where conflict is the signal of interest rather than a failure mode.
Bengs et al. prove that no strictly proper second-order scoring rule exists. In practice this means a network trained to output its own epistemic uncertainty by ordinary loss minimization is pushed toward degenerate answers. That result argues against the obvious design of training an evidential or credal head on camouflage labels. It is one reason we use a post-hoc readout that is never trained to report uncertainty.
We treat the readout's semantics as a hypothesis to test, not a guarantee. We do not claim it is a proper epistemic measure.
Research Program
Readout
We implement the readout for three off-the-shelf closed-world segmenters, all trained only on COD10K and CAMO in their original way:
- SINet, a two-stage search-then-identify network;
- CamoFormer, a transformer with masked separable attention;
- a DINOv2-guided model in the style of CODiff.
The readout costs one forward pass plus elementwise arithmetic on the head's inputs. Where the head is not a single linear map over nonnegative features, we apply the readout to the last such layer and report which layer was used.
Image-level summaries
We aggregate the pixel maps into two image-level numbers. The first is mean conflict inside the thresholded predicted mask. The second is mean ignorance over the whole image.
A three-way rule then assigns each image to camouflaged, salient, or background. Its thresholds are set only from quantiles on a held-out camouflaged-only validation split. For example, we reject as background when ignorance exceeds the 95th percentile seen on camouflaged images. This keeps the entire pipeline free of negative images.
We also test the pixel-level claim directly. Conflict should concentrate where foreground and background textures match, which we measure with local texture similarity across the ground-truth boundary. Ignorance should rise on scene types absent from training.
Supervised combination
In a second stage we ask whether the two maps help even when negatives are available. We feed them as extra input channels into OPCNet's existence module and retrain it on OPC16K.
Proposed Experiments
The main benchmark is OPC16K from Chen et al.. It has about 9,000 camouflaged images, about 3,000 pure backgrounds, and about 4,000 non-camouflaged-object images. We report the paper's own metrics:
- background false-positive rate ($FPR_{bg}$);
- false-positive rate on non-camouflaged-object images;
- three-way accuracy ($Acc_{all}$);
- camouflaged-object MAE ($coMAE$).
We add two threshold-free measures: AUROC for camouflaged versus background, and AUROC for camouflaged versus non-camouflaged objects.
Because the readout never changes the mask, closed-world segmentation quality on COD10K, CAMO, CHAMELEON and NC4K is unchanged by construction. We still report it after the supervised combination.
For the pixel-level ignorance test, we use DIOCblueberry as a visually camouflaged but out-of-domain scene type (agricultural foliage). There is no reported result found on Lacuna for any uncertainty method on DIOCblueberry, so this part is exploratory.
Baselines
Do-nothing reference. This is the closed-world segmenter itself, rejecting an image only when its thresholded mask is empty, or equivalently when the maximum pixel probability falls below a threshold. It anchors the failure mode: Chen et al. report negative-sample false-positive rates as high as 97.4% for RUN and USCNet. Any negative-free method must move far from this point to matter.
One-dimensional scores from the same forward pass. These are predictive entropy, max logit, and energy computed from the same model at the same threshold-calibration budget, meaning camouflaged-only validation quantiles. They are the cheapest competitors and see the same network. Beating them shows that the discarded sum-and-product dimension carries information the logit does not.
Sampling-based uncertainty. This is Monte Carlo dropout with 50 passes, matching UGTR, and a five-member deep ensemble of the same architecture, both trained on the same data. They are the standard strong answer for epistemic uncertainty. We report accuracy at equal wall-clock budget as well as at unlimited budget, so the readout is not favored only because it is cheap.
Learned error predictor and trained credal head. The first is a CANet-style error predictor (CANet). The second is a random-set head in the style of Random-Set Neural Networks, trained over the focal sets {O}, {B}, {O,B}. On ImageNet versus ImageNet-O, that random-set model reported AUROC 60.38 against 41.08 for Bayesian approximations. These two baselines test whether learning the uncertainty beats reading it out, which is the design the no-go theorem warns against.
Strongest existing method. This is OPCNet, trained with OPC16K negatives. It reports $FPR_{bg}$ 0.050, $Acc_{all}$ 0.905 and $coMAE$ 0.038, against $FPR_{bg}$ 0.076 for RUN and 0.179 for USCNet in the same table. The file reports both RUN's 0.076 and the 97.4% negative false-positive rate for RUN and USCNet. We read the latter as closed-world deployment and the former as the OPC16K comparison, and will confirm this against the paper.
Our negative-free readout is not expected to match OPCNet. We pre-register two success levels:
- Meaningful: beating USCNet's 0.179 $FPR_{bg}$ with zero negatives.
- Strong: reaching RUN's 0.076.
For the supervised combination, success means improving OPCNet's $Acc_{all}$ above 0.905 at the same training data and tuning budget.
Ablation of the key ingredient. This ablation replaces conflict and ignorance with the raw pair ($|z|$, $E^+ + E^-$). Dempster's rule is a monotone transform of these quantities. So the question is whether the belief-function form adds anything beyond "the head's total absolute contribution."
Basis-sensitivity control. This control recomputes $E^\pm$ after a random invertible remixing of the penultimate channels, folded into the next layer so the logits are unchanged. It checks whether the effect depends on the ReLU-defined channel basis the network actually uses.
Why not just…?
Why not just use a deep ensemble or Monte Carlo dropout? Ensemble disagreement measures whether models disagree about the logit, not whether each model sees conflicting evidence. On a pure background, five closed-world models can agree confidently on a hallucinated contour, because all were trained on the same always-positive data. The file offers indirect support: Random-Set Neural Networks found Bayesian approximations near chance on ImageNet-O, and EUQ reports sampling methods to be 100 to 1000 times slower. Still, ensembles are a real threat on the background split, and the equal-budget comparison settles it directly.
Why not just threshold entropy or max logit? These are functions of $E^+ - E^-$ alone, so by construction they cannot tell a blank patch from fur-on-leaf. If they nonetheless match our readout on OPC16K, the sum-and-product dimension carries no extra signal and the core claim fails. This comparison is therefore the first decisive check.
Why not just train an existence classifier on negatives, as OPCNet does? When negatives are available, this works well: 0.050 $FPR_{bg}$. We do not argue otherwise. The readout targets deployments where the negative distribution is unknown or shifts. Its second use is as input features, and the supervised-combination experiment tests whether it helps even the supervised system. If it adds nothing there, we will say its value is confined to the negative-free regime.
Why not just sum the absolute channel contributions and skip Dempster–Shafer? This is the most serious threat, because conflict and ignorance are monotone in the evidence totals. If the ($|z|$, $E^+ + E^-$) ablation performs identically, we will report that the finding is the two-dimensional evidence representation. Belief functions then serve as its interpretation, not as the source of the gain.
Decisive Checks
The idea stands or falls on three results:
- On OPC16K with zero negatives, conflict and ignorance must beat every one-dimensional score from the same network in both AUROCs.
- The two maps must behave differently on the two negative types. Ignorance should mainly separate backgrounds, and conflict should mainly separate non-camouflaged salient objects. If one map does all the work, the conflict/ignorance distinction is not real here.
- The effect must survive the basis-sensitivity control.
If all three hold across SINet, CamoFormer and the DINOv2-based model, we will have shown that the evidence needed to reject images with no camouflaged object is already inside closed-world detectors, discarded only by the sigmoid.
References
- Is There Really a Camouflaged Object? Towards Realistic Camouflaged Object Detection
- Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty Quantification
- Uncertainty-Guided Transformer Reasoning for Camouflaged Object Detection
- Predictive Uncertainty Estimation for Camouflaged Object Detection
- Uncertainty-Masked Bernoulli Diffusion for Camouflaged Object Detection Refinement
- Modeling Aleatoric Uncertainty for Camouflaged Object Detection
- On Second-Order Scoring Rules for Epistemic Uncertainty Quantification
- Camouflaged Object Detection
- CamoFormer: Masked Separable Attention for Camouflaged Object Detection
- CODiff: One-Step Diffusion Model for Camouflaged Object Detection
- Benchmarking Dense and Indiscernible Object Counting with Blueberries
- Random-Set Neural Networks