The Siamese-based object tracking algorithms employ random angle embedding to mitigate accuracy degradation caused by object rotation. However, it only adds prior knowledge as predefined patterns without alleviating the feature misalignment and local equivariance caused by rotation. In this paper, we propose a Siamese network with feature-calibration (AFC-Siam Net), which establishes the spatial correspondence relationship of the misalignment features for the same object under different rotation angles by dual-templates input structure with an affine transformation filter. Compared to existing Siamese-based trackers, AFC-Siam Net calibrates the features without the need for carefully designed prior knowledge. The experimental results confirm the effectiveness of the proposed algorithm through comprehensive evaluations on widely adopted datasets in the field of visual object tracking.