Search

Highlights • We formulate a new LRF with the projection by the scatter matrixs Eigen vectors and the mesh normal. • We firstly discuss the support radius for constructing the LRF and generating the feature descriptor separately. • In the experiments about Recall vs. 1-Precision Curve, we provide more precise measurements for TP (true positive) and FP (false positive). • Our methods concentrate on decreasing the calculation time and storage through all the processes. (e.g. low feature generation time and low feature dimension). Graphical abstract Abstract Object recognizing in cluttered scenes remains a largely unsolved problem, especially when applying feature matching to cluttered scenes there are many feature mismatches between the scenes and models. We propose our Feature Fusion Information Statistics (FFIS) as the calculation framework for extracting salient information from a Local Surface Patch (LSP) by a Local Reference Frame (LRF). Our LRF is defined on each LSP by projecting the scatter matrix's eigenvectors to a plane which is perpendicular to the normal of the LSP. Based on this, our FFIS descriptor of each LSP is calculated, for which we use the combined distribution of mesh and point information in a local domain. Finally, we evaluate the speed, robustness and descriptiveness of our FFIS with the state-of-the-art methods on several public benchmarks. Our experiments show that our FFIS is fast and obtains a more reliable matching rate than other approaches in cluttered situations. [ABSTRACT FROM AUTHOR]

Hypothesis Generation (HG) and Hypothesis Verification (HV) play an important role in 3D objection recognition. However, performing 3D object recognition in cluttered scenes using HG and HV still remains a largely unsolved problem. High False Positive (FP) in HG and HV stages are witnessed due to clutter and occlusion, which will further affect the final accuracy of recognition. To address these problems, we propose a novel Hough-space-based HG approach for extracting hypotheses. Differently from the existing methods, our approach is based on a Hough space which adopts a self-adapted measure to generate hypotheses. Based on this, a novel HV-based method is proposed to verify the hypotheses obtained from HG procedures. The proposed method is evaluated on four public benchmark datasets to verify its performance. Experiments show that our approach outperforms state-of-the-art methods, and obtains a higher recognition rate without sacrificing precision both at high FP rates and high occlusion rates. [ABSTRACT FROM AUTHOR]