Towards 3D VR-Sketch to 3D Shape Retrieval

Abstract

In this paper we emphasize the importance of unique certified one-time key pairs in Buyer-Seller Watermarking (BSW) protocols. We distinguish between reactive unbinding attacks, in which the seller reacts to illicit file sharing by fabricating further evidence of such activity, and pre-emptive unbinding attacks, in which the seller gains an advantage by taking action that pre-empts the file being shared. We demonstrate the importance of certified one-time key pairs in the BSW protocol by Lei et al., for protecting against pre-emptive unbinding attacks, and subsequently reveal a new attack on a recently published BSW protocol due to its omission of unique key pairs.

Fine-Grained VR Sketching: Dataset and Insights

“…3D sketch to 3D shape 3D sketch to 3D shape retrieval is a relatively novel field [38,37,80,20,47]. The progress is hampered by a lack of datasets of human 3D sketches.…”

“…Moreover, this dataset is no longer accessible online. Luo et al [47] addressed the category-level retrieval on sparse 3D VR sketches created by novices. They proposed a synthetic data generation method from 3D shapes and collected a small dataset of human sketches that they use as test data.…”

“…Given this drop, here we study how training on human sketches affects the retrieval performance, and how it compares with training on synthetic sketches. Unlike [47] we focus on the fine-grained retrieval and drive out insights on the desirable structure of 3D VR sketch datasets.…”

“…Research on free-hand VR sketches is hindered by a lack of training data. As a first step, Luo et al [47] collected a small dataset of 167 sketches for testing, and proposed a heuristic method to generate 3D sketches with different abstractness levels. They were the first to study the retrieval problem from sparse human VR sketches, and showed that a reasonable inter-class retrieval accuracy on human sketches can be achieved by training on synthetic sketches, even though inferior to the accuracy on synthetic data.…”

“…While in the absence of human sketches, synthetic sketches are the only option, the method for synthetic sketch generation proposed in [47] requires an input mesh to be a clean manifold curvature-aligned quad-dominant mesh. For new categories, collecting human sketches might be more feasible than obtaining 3D shapes with required properties.…”