PointCNN-Hand:利用PointCNN對手部點雲進行手部3D關節估測

Abstract

none

In this thesis, a novel method for 3D hand pose estimation entitled “PointCNN-Hand” is proposed. To effectively use the depth image, we transfer 2D depth images of the hand to 3D point cloud for estimation. We then feed the 3D point cloud into PointCNN-Hand to implement end-to-end training. Recently, hand joints estimation methods using depth images as inputs such as V2V, A2J, Hand-Pointnet, etc., have greatly reduced error rates. However, these researches have exposed a number of issues. To increase accuracy, the network computation is bound to be heavier. This is a critical issue, which makes real-time estimation of an implementation very difficult. In terms of network parameters, the proposed PointCNN-Hand has around one third of the parameter count of Hand-PointNet. Additionally, Hand-Pointnet requires principal components analysis (PCA) and an oriented bounding box (OBB) for the hand point cloud. With the proposed method, we need only to reposition the hand point cloud to the world coordinate center point, then reorient the values in the range [-1,1] via the MaxAbs (maximum absolute) standardization. To validate the performance of the proposed method, we perform error analysis on MSRA, NYU, ICVL datasets for comparison with the state-of-the-art methods. Experimental results show the proposed PointCNN-Hand has desired estimation results with fewer model parameters. The model works exceptionally well, while having a much smaller number of parameters of around 3MB, with a Floating-point operations (FLOPs) of only 232.05M.