Reliably detecting or tracking 3D features is challenging. It often requires preprocessing and filtering stages, along with fine-tuned heuristics for reliable detection. Alternatively, artificial intelligence-based strategies have recently been proposed; however, these typically require many manually labeled images for training. We introduce a method for 3D feature detection by using a convolutional neural network and a single 3D image obtained by fringe projection profilometry. We cast the problem of 3D feature detection as an unsupervised detection problem. Hence, the goal is to use a neural network that learns to detect specific features in 3D images using a single unlabeled image. Therefore, we implemented a deep-learning method that exploits inherent symmetries to detect objects with few training data and without ground truth. Subsequently, using a pyramid methodology of rescaling each image to be processed, we achieved feature detections of different sizes. Finally, we unified the detections using a Non-Maximum Suppression algorithm. Preliminary results show that the method provides reliable detection under different scenarios with a more flexible training procedure than other competing methods.