Classical methods of particle-image velocimetry (PIV) based on state-of-the-art cross-correlation algorithms constitute a spatial filtering operation that leads to a loss of spatial resolution of the estimated velocity fields. This is associated with numerous limitations, e.g., when it comes to aligning high-fidelity numerical simulations, since the detection of small turbulent scales is often precluded. An optical flow neural network adapted for PIV processing, called RAFT-PIV, can overcome the drawback of reduced spatial resolution while maintaining the reliability and robustness of the classical approach. The present work aims to extend the application of the aforementioned deep learning based method by integrating it within a stereoscopic PIV processing workflow. Experimental data from an optically accessible single-cylinder combustion engine the flow field of which is characterized by complex, three-dimensional flow structures and a broad range turbulent scales are processed using RAFT-PIV. The 2D-3C vector fields obtained from stereoscopic reconstruction are compared against a classical cross-correlation based PIV evaluation, providing insight into instantaneous quantities and turbulent structures. A qualitative comparison of the instantaneous in-plane and out-of-plane velocity components reveals that the reconstructed flow obtained from the deep learning based estimation includes a wide range of resolved turbulent length scales. It outperforms the classical method. Furthermore, the findings of the present study demonstrate that the pixel accuracy has an effect on both physical quantities and turbulence characteristics. In particular, an increase in turbulent kinetic energy is observed across the engine cycles of the investigated engine flow states. Moreover, a higher fraction of three-dimensional isotropic turbulent structures is derived from RAFT-PIV compared to the results from the cross-correlation based technique.