Intel Object Library - Training new objects

[PLajo1]

Hi!

The current Developer's Guide for the Intel Object Library does not explain how to train a classifier to recognize objects that are not included in the pre-trained classifier. In our application, we need to recognize and localize one specific object with Intel Realsense cameras.

1. Is it possible to train a classifier with the Object Library? How?

2. Otherwise, is there a recommended library to do custom object recognition and localization with Realsense cameras?

Thanks!

[Jeffrey_M_Intel1]

I am not an expert on realsense or the object library.  However, it looks like there are several paths to consider:

• The object library has a YOR tool.  "YOR is a tool that allows a developer or end user to train their own objects based on a number of snapshots of that object. This is a training tool that is paired with the Object Recognition API for Recognition."  https://software.intel.com/en-us/documentation/realsense-for-linux-developers-guide/yor-tool
• Point Cloud Library (PCL) works well with RealSense depth data, and has many options for object recognition.
• The state of the art for trainable 2D object recognition is deep learning training and inference.  There are several options for training with Intel optimized DL frameworks.  Intel's Inference Engine, in Computer Vision SDK or Deep Learning Deployment Toolkit, could be used with the 2D camera or 2D inference could be paired with PCL.

I will look for more info on the YOR tool and let you know if I find anything.