Continuous Perception Garment Classification Based on Optical Flow Variation

A novel continuous perception garment classification mechanism is proposed in this paper, with the aim to identify the correct category of the garment from a set of known categories. It has been observed that due to the severe folding and overlapped texture of garments, treating a video of the continuous deformation of cloth as a set of disordered static figures would be ineffective which leads to low classification precision performed by an image-based garment classifier. In contrast, a high-level decision making module that leverages the classification results of each single image would significantly increase the algorithm performance. In this paper, we incorporate the optical flow variation of deformable cloth between consecutive configurations as a representative of how it is traversing within the confidence interval of the image-based classifier. We claim that it is not the number of video frames but the sum of optical flow variation of the garment configuration between consecutive frames having the same category label that constitutes the belief of garment classification. In other words, if two consecutive visual appearances of the garment could be identified as the same category by the image-based classifier, then the more diverged that two configurations are, the more confident that the garment is correctly identified. Experimental comparisons between the state-of-the-art algorithm and the proposed algorithm in a public dataset have been provided which prove the validity of the proposed algorithm.