3D Human Mesh Regression With Dense Correspondence

3DBodyNet: Fast Reconstruction of 3D Animatable Human Body Shape From a Single Commodity Depth Camera

“…The experiments reveal that directly reconstructing the 3D human body shape from a single RGB image requires the body height as input due to the scale ambiguity from 2D to 3D. In contrast, depth images offer 3D information, which is exploited by [2], [38], [28] and our proposed method; this proves to us more accurate for 3D human body shape reconstruction compared to the use of a single 2D picture as in [47]. We also note that the training dataset used by [47] lacks RGB images of the human body in tight clothing which explains why [47] performs better when estimating the body shape under clothing than when reconstructing the body shape for tight clothing.…”

“…In contrast, depth images offer 3D information, which is exploited by [2], [38], [28] and our proposed method; this proves to us more accurate for 3D human body shape reconstruction compared to the use of a single 2D picture as in [47]. We also note that the training dataset used by [47] lacks RGB images of the human body in tight clothing which explains why [47] performs better when estimating the body shape under clothing than when reconstructing the body shape for tight clothing. Table IV compares our method with existing state-of-theart methods in terms of input data type, human body reconstruction, estimation of body shape under clothing, being animatable the speed and being deep learning.…”

“…Figure 8 and Table III compare our method and the related works for the task of estimating body shape under clothing. It is noticeable that the methods of [2], [38] and [47] do not perform well for predicting body shape under clothing. Compared to the results from Body PointNet [28], our results are better.…”

“…To quantitatively compare our algorithm with existing stateof-the-art methods, we perform experiments using the BUG dataset [28] as it has the ground-truth body shapes and dressed bodies. We compare our method against various state-of-theart methods including that of [2] that uses two depth-images as input, the single depth-image based method of [38], and the single RGB image based method of [47]. In contrast to these methods, our approach also works for estimating the body shape under clothing.…”

“…Compared to the results from Body PointNet [28], our results are better. As shown in Table II and Table III, it is important to observe that method [47] is a single RGB image-based method while the other approaches ( [2], [38], [28] and ours) are depth-based methods. The experiments reveal that directly reconstructing the 3D human body shape from a single RGB image requires the body height as input due to the scale ambiguity from 2D to 3D.…”