Lessons learned from lightweight CNN based object recognition for mobile robots

The focus of this paper is on the comparison of multiple neural network frameworks and the their usage in 2D/3D robot perception applications. Numerous frameworks exists for this task including the recent deep learning based ones, which allow us to develop a perception system, with the chosen parameters for object recognition. In this paper we analyzed the possible solutions, including different Convolutional Neural Networks (CNNs) variants. The advantages of 2D CNNs linked with 3D features lead to another approach, which can be extended further. The leading idea is to create a custom object recognition method that takes advantage of a 2D system's precision and speed, but it can efficiently incorporate 3D features. This way, the disturbances specific to each method separately can be minimized. On the other hand, this is a lightweight solution, that is supposed to be tolerated by less powerful processing units as well. By placing 3D bounding boxes around detected objects, the convenience of the 2D detection methods can be integrated in a 3D metric world.