In this paper, we explore the utility of classical neural network-based approaches, originally designed for processing 2D images, in semantic segmentation and object recognition tasks within the context of 3D point cloud images generated from handheld video recordings. Our investigation centers around the use of a custom-created, small-sized training dataset, consisting of 108 RGB images of humans and cobots in diverse industrial settings. This dataset allows us to demonstrate that flexible segmentation and recognition applications can be built even with a restricted dataset developed using widely available low-cost tools and modern convolutional neural net architectures. Downstream benefits of the approach include the ability to detect humans and human gestures, as well as to rapidly prototype digital twins in Industry 5.0 environments.