Summary
Smart Info-Media Systems in Asia
2019
Session Number:SS2
Session:
Number:SS2-4
3D Reconstruction from a Single Image Considering Connected Components
Takumi Suzuki, Tomoaki Kimura, Hiroyuki Tsuji,
pp.38-43
Publication Date:2019/9/4
Online ISSN:2188-5079
Summary:
Recently, the demand for 3D modeling methods has increased due to various technological developments, such as 3D printing and VR. However, handling 3D models manually is difficult; thus, a technology that automates 3D modeling is required. Currently, effective 3D reconstruction can be realized using disparity information; however, this method requires two or more images. Recently, reconstruction of 3D objects from a single image has been realized using an encoder-decoder deep learning model. However, the accuracy of the reconstructed objects is not sufficient because, in many cases, the reconstructed 3D objects may collapse. We propose a method to detect and reduce the collapse of reconstructed 3D objects by improving the deep learning model. The proposed model can detect the collapse by performing connected component labeling on the reconstructed 3D object. Then, the weight parameters are optimized to minimize the sum of the distances between connected components. As a result, compared to the conventional method, which causes overfitting and generates a 3D false shape, i.e., a shape that does not occur in the input image, the proposed method improves generalization performance. Experimental results quantitatively demonstrate that the accuracy rate of reconstructed voxels is improved; however, the recall rate is slightly lower than that of the conventional method. Moreover, the frequency of false object generation is reduced. In the future, a more effective metric to measure the collapse degree of reconstructed objects should be considered.