Post-training quantization (PTQ) is widely regarded as one of the most effective quantization techniques in practice. The current mainstream quantization strategy is generally to select appropriate quantization parameters by optimizing the block-wise reconstruction activation loss or global prediction difference, such as quantization scale factor and weight rounding parameter. However, these methods often neglect the impact of quantization errors on the attention distribution of critical discriminative regions in Class Activation Maps (CAMs). This may lead to a significant decrease in the classification performance of the quantization model at low bit-widths. To address this issue, this paper proposes a CAM-calibrated post-training quantization method (C-CAM). The core idea of this method is to design a differentiable CAM-aligned loss function that explicitly constrains the consistency in CAM space distributions between the full-precision and quantized models, thereby selecting optimal quantization parameters to enhance model performance. The experimental results on the ImageNet dataset demonstrate that C-CAM can not only be integrated into existing PTQ frameworks but also significantly outperforms current PTQ methods. For instance, it achieves a breakthrough by surpassing 20% classification accuracy for the first time on 2-bit quantized MobileNetV2.