Traditional Imitation Learning (IL) approaches often rely on teleoperation to collect training data, which ensures consistency between training and deployment action and observation spaces. However, teleoperation slows data acquisition, distorts expert behavior and data can be affected by the lack of teleoperation skills. To overcome these limitations, IL training on human demonstrations requires visual representations that are agnostic to both embodiment and environment. Recent advancements in Vision Foundation Models, such as Grounded-Segment-Anything (Grounded-SAM), offer a solution by extracting meaningful scene information while filtering out irrelevant details without manual annotation. In this work, we collected 50 human video demonstrations of a manipulation task from the RLBench benchmark. We evaluated Grounded-SAM's ability to automatically annotate objects of interest and proposed a 3D visual representation using depth maps. This representation was used to train a diffusion policy, which successfully generalized to simulated robot deployment in RLBench, despite being trained exclusively on real-world human demonstrations. Our results demonstrate that efficient training can be achieved with just 50 demonstrations and halfan-hour training time.