Grasping Field: Learning Implicit Representations for Human Grasps

Abstract

Figure 1: Ground truth grasps and generated grasps. Each row corresponds to one object. Left three columns show the ground truth grasps, each from three different viewpoints. The middle three columns show one generated example, and the right three columns show another generated example. Note that these objects are never seen during training. See Appendix E (Fig. E.4) for more examples.

LEAP: Learning Articulated Occupancy of People

“…They are traditionally modeled either as linear combinations of analytic functions or as signed distance grids, which are flexible but memory expensive [55]. Even though the problem of the memory complexity for the grid-based methods is approached by [27,43,57,66,67], they have been outperformed by the recent learning-based continuous representations [2,3,10,12,19,30,39,40,42,45,46,56,62,64]. Furthermore, to improve scalability and representation power, the idea of using local features has been explored in [7,11,41,46,51,52,62].…”

A Skeleton-Driven Neural Occupancy Representation for Articulated Hands

“…To the best of our knowledge, there are no implicit hand representations that can generalize well to various shapes. Grasping Field [29] learns an implicit function for hand and objects together to represent contact but treats every posed hand as a rigid object. As a result, the complexity of learning a wide range of poses increases significantly.…”

“…Hand-object interaction. There have been many studies into hand interaction with objects in various settings [5,7,9,12,17,18,23,24,29,31,42,59]. Recently, the community has begun exploring the task of generating plausible hand grasps given an object with notable studies including [12], [29], and [59].…”

“…There have been many studies into hand interaction with objects in various settings [5,7,9,12,17,18,23,24,29,31,42,59]. Recently, the community has begun exploring the task of generating plausible hand grasps given an object with notable studies including [12], [29], and [59]. GanHand [12] generates grasps for each object in a given RGB image by first predicting a grasp type from grasp taxonomy [17] and its initial orientation, and then optimizing for better contact with the object.…”

“…The predicted hand is then fed to a refinement model to improve the contact. Grasping Field [29] learns a signed distance field for both hand surface and object surface in one space, allowing the contact to be learned as regions where distances to both surfaces are zeros. However, the output surface cannot be x c articulated and requires hand model fitting.…”