Resumé
This paper reports the discovery of a fast, yet highly discriminative local 3D descriptor for point cloud data. Local descriptors are popular and highly effective for various 3D tasks such as registration, pose estimation and object recognition. Good solutions for these tasks critically depend on the ability to make correct associations between two or more models, or the local features on these, even under the influence of disturbances such as noise, clutter and occlusions. Our descriptor formulation is inspired by the geometric relations employed by the well-known Point Pair Feature, used originally on a global scale for classification and later on a semi-global scale for recognition. We have identified the most discriminative subset of relations for use in a local descriptor, resulting in a condensed representation of the local variation around a surface point. We compare against seven competing mesh and point cloud descriptors on eight different matching benchmarks with a well-defined evaluation protocol. In all cases, our descriptor outperforms earlier works, providing relative gains in accuracy above 100 % for two of the four real datasets considered. Finally, we subject all descriptors to RANSAC based pose estimation and object recognition evaluation on four real datasets. In all four cases, our descriptor matches or surpasses state of the art performances.
| Originalsprog | Engelsk |
|---|---|
| Publikationsdato | 6. sep. 2018 |
| Status | Udgivet - 6. sep. 2018 |
| Begivenhed | 29th British Machine Vision Conference, BMVC 2018 - Newcastle, Storbritannien Varighed: 3. sep. 2018 → 6. sep. 2018 |
Konference
| Konference | 29th British Machine Vision Conference, BMVC 2018 |
|---|---|
| Land | Storbritannien |
| By | Newcastle |
| Periode | 03/09/2018 → 06/09/2018 |
| Sponsor | Amazon, et al., Microsoft, NVIDIA, SCANs, SCAPE |
Fingeraftryk
Citer dette
}
Local point pair feature histogram for accurate 3D matching. / Buch, Anders Glent; Kraft, Dirk.
2018. Afhandling præsenteret på 29th British Machine Vision Conference, BMVC 2018, Newcastle, Storbritannien.Publikation: Konferencebidrag uden forlag/tidsskrift › Paper › Forskning › peer review
TY - CONF
T1 - Local point pair feature histogram for accurate 3D matching
AU - Buch, Anders Glent
AU - Kraft, Dirk
PY - 2018/9/6
Y1 - 2018/9/6
N2 - This paper reports the discovery of a fast, yet highly discriminative local 3D descriptor for point cloud data. Local descriptors are popular and highly effective for various 3D tasks such as registration, pose estimation and object recognition. Good solutions for these tasks critically depend on the ability to make correct associations between two or more models, or the local features on these, even under the influence of disturbances such as noise, clutter and occlusions. Our descriptor formulation is inspired by the geometric relations employed by the well-known Point Pair Feature, used originally on a global scale for classification and later on a semi-global scale for recognition. We have identified the most discriminative subset of relations for use in a local descriptor, resulting in a condensed representation of the local variation around a surface point. We compare against seven competing mesh and point cloud descriptors on eight different matching benchmarks with a well-defined evaluation protocol. In all cases, our descriptor outperforms earlier works, providing relative gains in accuracy above 100 % for two of the four real datasets considered. Finally, we subject all descriptors to RANSAC based pose estimation and object recognition evaluation on four real datasets. In all four cases, our descriptor matches or surpasses state of the art performances.
AB - This paper reports the discovery of a fast, yet highly discriminative local 3D descriptor for point cloud data. Local descriptors are popular and highly effective for various 3D tasks such as registration, pose estimation and object recognition. Good solutions for these tasks critically depend on the ability to make correct associations between two or more models, or the local features on these, even under the influence of disturbances such as noise, clutter and occlusions. Our descriptor formulation is inspired by the geometric relations employed by the well-known Point Pair Feature, used originally on a global scale for classification and later on a semi-global scale for recognition. We have identified the most discriminative subset of relations for use in a local descriptor, resulting in a condensed representation of the local variation around a surface point. We compare against seven competing mesh and point cloud descriptors on eight different matching benchmarks with a well-defined evaluation protocol. In all cases, our descriptor outperforms earlier works, providing relative gains in accuracy above 100 % for two of the four real datasets considered. Finally, we subject all descriptors to RANSAC based pose estimation and object recognition evaluation on four real datasets. In all four cases, our descriptor matches or surpasses state of the art performances.
UR - http://www.scopus.com/inward/record.url?scp=85072347491&partnerID=8YFLogxK
M3 - Paper
ER -