When W4A4 Breaks Camouflaged Object Detection: Token-Group Dual-Constraint Activation Quantization

Camouflaged object detection (COD) segments objects thatintentionally blend with the background, so predictions depend on subtletexture and boundary cues. COD is often needed under tight on-devicememory and latency budgets, making low-bit inference highly desirable.However, COD is unusually hard to quantize aggressively. We studypost-training W4A4 quantization of Transformer-based COD and finda task-specific cliff: heavy-tailed background tokens dominate a sharedactivation range, inflating the step size and pushing weak-but-structuredboundary cues into the zero bin. This exposes a token-local bottleneck—remove cross-token range domination and bound the zero-bin mass under4-bit activations. To address this, we introduce COD-TDQ, a COD-aware Token-group Dual-constraint activation Quantization method.COD-TDQ addresses this token-local bottleneck with two coupled steps:Direct-Sum Token-Group (DSTG) assigns token-group scales to sup-press cross-token range domination, and Dual-Constraint Range Projection(DCRP) projects each token-group clip range to keep the step-to-dispersionratio and the zero-bin mass bounded. Across four COD benchmarksand two baseline models (CFRN and ESCNet), COD-TDQ consistentlyachieves an Sα score more than 0.12 higher than that of the state-of-the-art quantization method without retraining. The code is available athttps://github.com/MCG-NKU/nku-model-compre.