Recent high-contrast imaging surveys, looking for planets in young, nearby systems showed evidence of a small number of giant planets at relatively large separation beyond typically 20 au where those surveys are the most sensitive. Access to smaller physical separations between 5 and 20 au is the next step for future planet imagers on 10 m telescopes and ELTs in order to bridge the gap with indirect techniques (radial velocity, transit, astrometry with Gaia). In that context, we recently proposed a new algorithm, Keplerian-Stacker, combining multiple observations acquired at different epochs and taking into account the orbital motion of a potential planet present in the images to boost the ultimate detection limit. We showed that this algorithm is able to find planets in time series of simulated images of SPHERE even when a planet remains undetected at one epoch. Here, we validate the K-Stacker algorithm performances on real SPHERE datasets, to demonstrate its resilience to instrumental speckles and the gain offered in terms of true detection. This will motivate future dedicated multi-epoch observation campaigns in high-contrast imaging to search for planets in emitted and reflected light. Results. We show that K-Stacker achieves high success rate when the SNR of the planet in the stacked image reaches 7. The improvement of the SNR ratio goes as the square root of the total exposure time. During the blind test and the redetection of HD 95086 b, and betaPic b, we highlight the ability of K-Stacker to find orbital solutions consistent with the ones derived by the state of the art MCMC orbital fitting techniques, confirming that in addition to the detection gain, K-Stacker offers the opportunity to characterize the most probable orbital solutions of the exoplanets recovered at low signal to noise.