Concealment-Aware Verdicts for Claims About Hidden Objects in Images

Many viral claims about photographs are claims about things that are hard to see: "there is a sniper in this tree line," "a snow leopard is hiding on this rock face," "a vehicle is concealed under that netting." A fact-checker answering such a claim faces a problem that text-only verification never meets. If a careful look finds nothing, that does not show the claim is false, because the object may be well hidden. Today's systems ignore this. Camouflaged object detection, the vision task of segmenting objects that match their background's color and texture, assumes the object is always there. Multimodal fact-checkers use the image as a search query or ask a general vision-language model, and either one will say "no" whenever it sees nothing.

We propose a claim verifier for hidden-object claims. It returns one of three verdicts: Supported (target found), Refuted (target absent, and the image is clear enough that it would have been found), or Not Enough Info (nothing found, but the scene could plausibly hide it). Each verdict comes with a visual justification: a segmentation mask for Supported, or a map of the regions where the target could still be hiding for Not Enough Info. The core move is to split the verdict into two learned quantities. One is how strongly the image shows the target. The other is how likely the verifier is to miss the target in this particular scene. These are combined with an explicit prior on presence, so the system never treats "I didn't see it" as "it isn't there" in cluttered, high-concealment scenes.

Core Claim

Our claim is that, for presence claims about concealed targets, an estimated per-image miss rate improves verdicts in three ways. It reduces false refutations, meaning cases where the system calls a claim false while the target is actually present. It is robust when the share of true claims changes between training and deployment. It also yields explanations that a human can audit. A single calibrated presence classifier does not offer all three. The decision rule is plain Bayes. Let $\pi$ be the prior probability that the claimed target is present. Let $m$ be the estimated probability that the segmenter misses a present target in this image, and $f$ the probability that it fires on an absent one. Then, after a non-detection,

$$P(\text{present}\mid\text{no detection}) = \frac{\pi\, m}{\pi\, m + (1-\pi)(1-f)}.$$

The system says Refuted only when this posterior is below a threshold $\tau_R$. It says Not Enough Info when the posterior lies between $\tau_R$ and a detection threshold. Because $\pi$ is an explicit input rather than something absorbed from the training data, it can be set per deployment. Real claim streams have unknown and shifting base rates, and an adversary can choose them.

This differs from the closest work in specific ways. Referring Camouflaged Object Detection (R2CNet) segments a specific camouflaged category, but it is guided only by images of salient examples, it assumes the target exists, and it outputs a mask rather than a verdict. Is There Really a Camouflaged Object? questions the existence assumption but does not tie it to claims or abstention. LOREN supplies the three-valued, phrase-level verdict logic and "culprit" detection we borrow, but it is purely textual. Its rule "Refuted if any phrase is refuted" has no notion that evidence can be missing because it is hidden. DEFAME and How to Train Your Fact Verifier handle multimodal claims, but they treat the image globally through retrieval or CLIP/LLaVA-style encoders. Camouflage work such as Camouflaged Object Detection suggests such encoders are poorly suited to concealed boundaries.

We considered and rejected the most obvious idea at this intersection: using camouflage detectors to find seamless diffusion inpainting in news images. CSI-IMD already shows that forensic models trained for this, TruFor and CAT-Net v2, reach IoU 0.652 and 0.614 on such edits. A reviewer would rightly say "fine-tune TruFor on the 500,000-image CSI-IMD Extended set." That approach exploits generator traces that camouflage models are not built to see. We think it would likely match a camouflage-based detector, so we do not pursue it.

Research Program

The work has three parts. First, a benchmark of hidden-object claims. Second, a text-conditioned camouflage verifier with a miss-rate head. Third, a phrase-level explanation layer.

Benchmark. No benchmark in the file contains claims about concealed targets, so we build one from existing camouflage data. COD10K has 10,000 images over 78 categories with object-level masks and attribute tags such as "Small Object" and "Occlusion." R2C7K has 7,000 camouflaged images over 64 categories. The background-only images used in the existence-aware COD work supply true-absent cases. Each image receives claims of the form "a [category] is hidden [in region]."

The claims fall into three groups. True claims name the annotated category. False-category claims name a different, visually plausible category. False-presence claims are attached to images with no object. Region phrases, such as "on the left branch," are derived from the masks. This gives LOREN-style compound claims, where the wrong phrase can be either the entity or the location. Templates are paraphrased by a language model and spot-checked by hand.

We also build a "hard-present" split. These are present targets that a panel of human viewers fails to find within a fixed time budget, so Not Enough Info is a defensible answer for them. This split gives us human ground truth for when abstention is right. We will also collect a small real-world set of viral "hidden X" claims as a qualitative stress test. It will be reported separately and is too small for headline numbers.

Verifier. The segmentation backbone follows CamoFormer. Its attention is split into heads restricted to foreground, background, and the whole image. We add a text query: the embedded entity phrase cross-attends into the foreground and background heads, playing the role that R2CNet's salient-example branch plays. The model produces three outputs from shared features: a presence score $s$, an estimated miss rate $m$, and an estimated false-fire rate $f$.

The miss-rate head needs no extra annotation. On held-out folds, we record whether the segmenter's own prediction overlaps a present ground-truth target. The head is trained to predict that outcome from image features and from the uncertainty map produced by an uncertainty-guided decoder, following Uncertainty-Guided Transformer Reasoning. Since the head learns to predict the segmenter's own failures, $m$ is high exactly where texture matching and suppressed edges defeat the segmenter. We train $f$ the same way on absent images.

Explanation layer. Compound claims are split into entity and location phrases, as in LOREN. Each phrase gets its own three-valued verdict, and the claim verdict is aggregated by LOREN's rule with one change. A phrase whose verdict is Not Enough Info because of concealment cannot produce a claim-level Refuted. The explanation has three parts: the mask, or the high-$m$ regions when nothing is found; the verdict for each phrase; and a one-sentence rendering of the posterior, written from a template so it cannot add claims the model did not compute.

Baselines

The naive reference point is a verifier that always outputs Supported, together with a presence-only rule that says Refuted whenever the segmenter finds nothing. The second is the behavior we expect from current systems, and it fixes the false-refutation rate we aim to reduce.

The strongest existing method for locating a named camouflaged category is R2CNet. To feed it a text claim, we retrieve K salient images of the claimed category from the R2C7K Ref-subset. Referring Camouflaged Object Detection reports weighted F-measure 0.669 on R2C7K for R2CNet, against 0.604 for a plain COD baseline, with K=5 working best. We use K=5 and threshold its mask area for presence. For segmentation quality in the general setting, CamoFormer reports weighted F-measure 0.786 on COD10K (CamoFormer). Our text-conditioned model must stay close to that number on present targets, or a verdict gain could simply come from worse masks.

The second strong competitor is a two-stage pipeline. Class-agnostic CamoFormer proposes camouflaged regions, and a CLIP classifier scores each crop against the claimed category, with the presence threshold tuned on validation data. A third is a vision-language model (GPT-4o or LLaVA) prompted with chain-of-thought to find the target and give a three-way verdict. No reported result was found on Lacuna for vision-language models on camouflaged-object claim verification. The nearest figure is vanilla GPT-4o's 35.2% accuracy on post-cutoff ClaimReview2024+ in DEFAME, which concerns a different task. No reported result was found on Lacuna for the existence-aware COD model of Is There Really a Camouflaged Object? either. We will retrain it on our splits as a presence detector.

The key ablation replaces the decomposed posterior with a single presence classifier trained on the same images and features, calibrated by temperature scaling, and thresholded into three verdicts. This isolates whether the separate miss-rate head and the explicit prior add anything.

All trained methods share the same backbone and training images, and each gets the same hyperparameter search budget on a shared validation split. Every method's three-way thresholds are tuned to the same target coverage. We count a method as beaten only on two conditions. Our verifier must have a lower false-refutation rate at matched coverage and matched Supported precision. And it must keep that advantage when the test-time presence rate moves away from the training rate.

Why not just…?

Why not just prompt a vision-language model to look carefully and answer? The evidence suggests these models describe images globally and confabulate about specific pixels. VizAnchor reports intent-reasoning accuracy of 0.44 for an unanchored baseline against 0.86 for its pixel-anchored system. Nothing in the file, however, measures vision-language models on camouflaged targets directly. We therefore run them as a full baseline on both the hard-present and true-absent splits. If they abstain appropriately on hard-present images, the miss-rate machinery is unnecessary and we will say so.

Why not just run a camouflage detector and classify the crop with CLIP? This is a serious competitor for Supported verdicts, and it may match us there. Its weakness is the other two verdicts. When the class-agnostic detector finds nothing, it either hallucinates a mask, since COD training assumes presence, or returns nothing. In either case the pipeline has no basis for choosing Refuted over Not Enough Info. NewsCLIPpings also shows CLIP is fooled by semantically plausible distractors, which our false-category claims deliberately supply. The hard-present and prior-shift experiments decide this comparison.

Why not just turn text into example images and use R2CNet? Retrieving examples is simple and works for categories in R2C7K. It does nothing about the assumption that the target exists, and its accuracy is limited by the gap between salient and camouflaged appearance that the R2CNet paper itself identifies. We include it at full strength. If it matches our segmenter's masks, we keep it as the segmentation backbone and attach our miss-rate head to it. The contribution is the verdict rule, not the encoder.

Why not just calibrate one presence classifier? This is the most serious threat. A well-calibrated $P(\text{present}\mid\text{image})$ already contains the effect of concealment, so with the training base rate it could match our verdicts. Our case for the decomposition rests on two points. First, the prior should be adjustable at deployment, which a single calibrated score cannot do cleanly. Second, the separate miss-rate estimate lets the explanation say why the system abstained. If the calibrated classifier matches us under prior shift once a standard label-shift correction is applied, the contribution shrinks to the benchmark and the explanation format. The ablation is designed to show which case holds.

Proposed Experiments

The headline experiment reports, on the combined COD10K/R2C7K claim benchmark, three-way accuracy, macro-F1, and false-refutation rate at fixed coverage. Results are broken down by COD10K attribute tags, which shows whether the gains concentrate on occluded and small targets as the mechanism predicts.

The decisive check is prior shift. Test sets are rebuilt with 10%, 50%, and 90% of claims true, while all models keep their training-time thresholds. Our verifier receives $\pi$, and the calibrated-classifier ablation receives a label-shift correction. A decomposition that actually helps should keep its false-refutation rate flat across these settings. It should also beat the ablation by a margin larger than the spread across random seeds.

On the hard-present split, we measure agreement with human abstention. The question is whether Not Enough Info verdicts fall on the images people also failed to search successfully, rather than on random ones.

For compound claims, we report culprit F1, the share of false claims where the system blames the right phrase, as defined in LOREN. LOREN's 74.3 on textual FEVER claims is a reference point only, not a comparable target.

Explanation faithfulness gets two tests. A deletion test fills in the predicted mask with a background-matching inpainter and checks that the Supported verdict flips. A small study then asks fact-checkers whether the high-miss-rate region maps help them decide where to look.

Finally, we test whether the miss-rate head transfers to dense clutter on DIOCblueberry, using claims of the form "at least N berries are visible." Here a miss is common by construction, so the posterior should push the system toward abstention rather than confident undercounting.