Abstract
3D edge features, which represent the boundaries between different objects or surfaces in a 3D scene, are crucial for many computer vision tasks, including object recognition, tracking, and segmentation. They also have numerous real-world applications in the field of robotics, such as vision-guided grasping and manipulation of objects. To extract these features in the noisy real-world depth data, reliable 3D edge detectors are indispensable. However, currently available 3D edge detection methods are either highly parameterized or require ground truth labelling, which makes them challenging to use for practical applications. To this extent, we present a new 3D edge detection approach using unsupervised classification. Our method learns features from depth maps at three different scales using an encoder-decoder network, from which edge-specific features are extracted. These edge features are then clustered using learning to classify each point as an edge or not. The proposed method has two key benefits. First, it eliminates the need for manual fine-tuning of data-specific hyper-parameters and automatically selects threshold values for edge classification. Second, the method does not require any labelled training data, unlike many state-of-the-art methods that require supervised training with extensive hand-labelled datasets. The proposed method is evaluated on five benchmark datasets with single and multi-object scenes, and compared with four state-of-the-art edge detection methods from the literature. Results demonstrate that the proposed method achieves competitive performance, despite not using any labelled data or relying on hand-tuning of key parameters.