Pixel-wise hand segmentation of multi-modal hand activity video dataset

A method for generating a multi-modal video dataset with pixel-wise hand segmentation is disclosed. To address the challenges of conventional dataset creation, the method advantageously utilizes multi-modal image data that includes thermal images of the hands, which enables efficient pixel-wise hand segmentation of the image data. By using the thermal images, the method is not affected by fingertip and joint occlusions and does not require hand pose ground truth. Accordingly, the method can produce more accurate pixel-wise hand segmentation in an automated manner, with less human effort. The method can thus be utilized to generate a large multi-modal hand activity video dataset having hand segmentation labels, which is useful for training machine learning models, such as deep neural networks.