Robust, domain-invariant pose transformation remains a central challenge in building reliable gesture-based interfaces. One of the main limitations of current systems is their sensitivity to variations in camera viewpoint and signer orientation, which often require large-scale, multi-view datasets to ensure generalization. Collecting and annotating such datasets can be costly and impractical, especially for applications that aim to operate in diverse, unconstrained environments such as smart homes or ambient intelligence systems. This paper explores a modular approach to address this issue by introducing an intermediate transformation stage that translates pose coordinates with respect to viewpoint. Specifically, we propose a synchronised multi-camera setup during training, in which a dedicated module based on a convolutional autoencoder with skip connections learns to adapt skeletal data captured from arbitrary camera angles into a consistent, front-facing representation. The network jointly processes normalized 2D keypoints and their absolute positions, merging these features in the bottleneck to produce heatmaps of the transformed pose. Once trained, this module enables inference from a single arbitrary viewpoint by projecting the observed pose into the canonical view space. By decoupling viewpoint adaptation from downstream tasks, new viewpoints can be incorporated without retraining the core models, simply by updating the transformation component. We evaluate our approach using two newly recorded datasets captured simultaneously from frontal and top-down cameras. Preliminary results show that the proposed autoencoder effectively transforms poses from a top-view perspective into a canonical frontal representation, achieving low reprojection errors even on unseen pose variations.