Object detection and recognition are central problems in computer vision literature and essential for a vision based library. Recent advances in Covolutional Neural Networks (CNN)s have made the detection have made the recognition problem tractable for large number of object categories that would have been very expensive with model based classification approaches. In this project I will implement state-of-art object recognition as well as object detection/localization technique for RoboComp library. The implementation will support both real images and rendered images from CAD model. The implementation will be based on CUDA library and allow user to train and test his models. Also, a caffe independent pure RoboComp based implementation for forward pass will be developed for a selected CNN. Another component based on open scene graph will designed that would allow reading of CAD models and interfacing with current RoboComp simulation framework. Selection of rich textured CAD models is an crucial for any detection algorithm. Therefore, the proposed project will accompany a dataset of cleaned and textured CAD models for 5 object categories.





  • Rajath Kumar M P
  • marcog