CFCamo framework teaches AI when to abstain from detecting camouflaged objects

Camouflaged object detection (COD) systems are typically trained only on images that contain a target, which creates an over-detection bias: they often hallucinate objects in ordinary images that lack any camouflage. To address this, the authors of CFCamo first construct Counterfactual COD (CF-COD), a paired benchmark where the camouflaged target is digitally removed from each evaluation image while preserving a plausible background. This benchmark measures whether a model correctly detects the target on the original and abstains on the target-absent counterfactual, quantified by a new metric called Pair Accuracy (PA).

Their proposed framework, CFCamo, trains a Qwen3-VL-4B-Instruct agent using Counterfactual Sequence Policy Optimization (CSPO). CSPO samples paired rollouts of original and counterfactual images and applies a Counterfactual Paired Reward (CPR) that couples detection on the original with abstention on the counterfactual. On standard benchmarks, CFCamo improves S_alpha by 3.7 percentage points over prior RL-based COD methods. On CF-COD, it reaches 80.0-90.8% PA, while ablations show that removing counterfactual coupling collapses PA to just 1.4-5.2%, despite retaining strong target-present scores. This confirms that traditional evaluation alone fails to characterize reliable detect-or-abstain behavior.