Joint spatial-depth feature pooling for RGB-D object classification

TY - GEN

T1 - Joint spatial-depth feature pooling for RGB-D object classification

AU - Pan, Hong

AU - Olsen, Søren Ingvor

AU - Zhu, Yaping

N1 - Conference code: 19

PY - 2015

Y1 - 2015

N2 - RGB-D camera can provide effective support with additional depth cue for many RGB-D perception tasks beyond traditional RGB information. However, current feature representations based on RGB-D camera utilize depth information only to extract local features, without considering it for the improvement of robustness and discriminability of the feature representation by merging depth cues into feature pooling. Spatial pyramid model (SPM) has become the standard protocol to split 2D image plane into sub-regions for feature pooling in RGB-D object classification. We argue that SPM may not be the optimal pooling scheme for RGB-D images, as it only pools features spatially and completely discards the depth topological information. Instead, we propose a novel joint spatial-depth pooling scheme (JSDP) which further partitions SPM using the depth cue and pools features simultaneously in 2D image plane and the depth direction. Embedding the JSDP with the standard feature extraction and feature encoding modules, we achieve superior performance to the state-ofthe- art methods on benchmarks for RGB-D object classification and detection.

AB - RGB-D camera can provide effective support with additional depth cue for many RGB-D perception tasks beyond traditional RGB information. However, current feature representations based on RGB-D camera utilize depth information only to extract local features, without considering it for the improvement of robustness and discriminability of the feature representation by merging depth cues into feature pooling. Spatial pyramid model (SPM) has become the standard protocol to split 2D image plane into sub-regions for feature pooling in RGB-D object classification. We argue that SPM may not be the optimal pooling scheme for RGB-D images, as it only pools features spatially and completely discards the depth topological information. Instead, we propose a novel joint spatial-depth pooling scheme (JSDP) which further partitions SPM using the depth cue and pools features simultaneously in 2D image plane and the depth direction. Embedding the JSDP with the standard feature extraction and feature encoding modules, we achieve superior performance to the state-ofthe- art methods on benchmarks for RGB-D object classification and detection.

U2 - 10.1007/978-3-319-19665-7_26

DO - 10.1007/978-3-319-19665-7_26

M3 - Article in proceedings

SN - 978-3-319-19664-0

T3 - Lecture notes in computer science

SP - 314

EP - 326

BT - Image analysis

A2 - Paulsen, Rasmus R.

A2 - Pedersen, Kim S.

PB - Springer

T2 - Scandinavian Conference, SCIA 2015

Y2 - 15 June 2015 through 17 June 2015

ER -