Depth Kernel Descriptors for Object Recognition

This paper introduces their devised depth image feature descriptor. Then, it compares with previous feature descriptors. It suggests results as depth only, rgb only and depth/rgb. As it used open image data, I can compare my result with this. 

(http://rgbd-dataset.cs.washington.edu/)

The experiment happended in instance recognition and category recognition

The impressive statements are like below. 

1. Introduction

we present five depth kernel descriptors that capture different recognition cues including size, shape and edges (depth discontinuities). Extensive experiments suggest that these novel features complement each other and their combination significantly boosts the accuracy of object recognition on an RGB-D object dataset as compared to state-of-the-art techniques. 

2. Kernel descriptors over depth maps

- Size features over Point Clouds

- Shape Features over Point Clouds

- Edge Features over Depth Maps

4. Experiments

We distinguish between two types of object recognition: instance recognition and category recognition. 

A. Dataset

We evaluate depth kernel descriptors on a large-scale multi-view object dataset collected using an RGB-D camera (http://rgbd-dataset.cs.washington.edu/)