Color-Assisted Local Feature Pipeline for Three-Dimensional Object Retrieval

We propose a practical color-assisted local feature-based pipeline for 3D object instance recognition with a fixed imaging system which poses certain constrains for object viewing angle and distance. Assuming the objects are rigid and are put on a flat surface when being recognized, our approaches take advantages of the known system setup to reduce the number of samples needed for training, while achieving faster recognition with higher accuracy than general purpose methods. The system is first calibrated for object segmentation and white balancing, which removes background points and generate reliable object color information. Then we modify the traditional local feature pipeline for 3D object recognition by adding color similarity constrains both at the candidate object level and at the local feature points level. Experimental results show significant improvement on recognition accuracy. A demonstration system is created to show the real-time performance of our methods.