Segmentation and semantic labelling of RGBD data with convolutional neural networks and surface fitting

摘要: We present an approach for segmentation and semantic labelling of RGBD data exploiting together geometrical cues deep learning techniques. An initial over-segmentation is performed using spectral clustering a set non-uniform rational B-spline surfaces fitted on the extracted segments. Then convolutional neural network (CNN) receives in input colour geometry with surface fitting parameters. The made nine stages followed by softmax classifier produces vector descriptors each sample. In next step, iterative merging algorithm recombines output into larger regions matching various elements scene. couples adjacent segments higher similarity according to CNN features are candidate be merged accuracy used detect which belong same surface. Finally, labelled obtained combining from CNN. Experimental results show how proposed outperforms state-of-the-art methods provides accurate labelling.