Semantic-only Visual Odometry based on dense class-level segmentation

Abstract : This paper proposes a novel approach called Semantic Visual Odometry (SemVO) which incorporates class-level consistency priors into the problem of 6-DoF Visual Odometry. Dense class-level labels are learnt for each pixel of the image using a CNN trained for semantic segmentation. A semantic error is formulated penalising the sum of squared differences (SSD) on class-level feature maps extracted from the decoder of a RefineNet. It will be shown how the proposed approach allows dense RGB-D camera tracking using solely a semantic error term. SemVO is evaluated on the ScanNet dataset and the results demonstrate how the number of classes affects performance. Results are also provided showing how best to fuse the new error function with classic dense photometric and geometric methods. Finally, it is demonstrated that SemVO improves over standard approaches for large camera motion applications.
Liste complète des métadonnées

https://hal.archives-ouvertes.fr/hal-01874544
Contributor : Andrew Comport <>
Submitted on : Friday, September 14, 2018 - 2:00:02 PM
Last modification on : Thursday, February 7, 2019 - 5:48:32 PM
Document(s) archivé(s) le : Saturday, December 15, 2018 - 2:08:49 PM

File

Howard_MAHE_ICPR18_0741.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01874544, version 1

Collections

Citation

Howard Mahe, Denis Marraud, Andrew I. Comport. Semantic-only Visual Odometry based on dense class-level segmentation. International Conference on Pattern Recognition (ICPR 2018), Aug 2018, Pékin, China. ⟨hal-01874544⟩

Share

Metrics

Record views

93

Files downloads

263