An Invariant Tangent-Angle Descriptor and a Band U-Net for 2D Fragment Adjacency Prediction
Guillaume Brouillette, Alain Goupil, Pierre-Olivier Parisé, Fadel Touré
原文摘要
Reassembling a broken artefact begins with deciding which fragments are adjacent. This paper addresses adjacency prediction for pairs of 2D fragments by analysing their contours. We propose a two-stage pipeline. In the first stage, each contour point is encoded using a tangent-angle profile, a descriptor that does not depend on rotation, translation or the choice of starting point. Pairs of profiles are then used to build an adjacency matrix, each entry computed either from a log-likelihood ratio or using a small 1D convolutional network. In the second stage, a U-Net classifies this matrix to predict whether the two fragments are adjacent. We compared our pipeline with two approaches from the literature using their own datasets. The results show that our approach outperforms them by reaching a Recall@10 of 0.823 versus 0.564 on the real dataset of PairingNet, and an accuracy of 0.979 versus 0.916 on Beaulac's dataset.
cs.CV · cs.LG