2021/08/21 by Issam Laradji, Laradji, Issam, Pau Rodríguez +5
Computer Science · Earth and Planetary Sciences · Engineering · #3D Shape Modeling and Analysis #3D Surveying and Cultural Heritage #Advanced Vision and Imaging #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences
paper · pdf · doi:10.48550/arxiv.2108.09593
openalex publication_date 2021/08/21 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Recent work has made significant progress in learning object meshes with weak\nsupervision. Soft Rasterization methods have achieved accurate 3D\nreconstruction from 2D images with viewpoint supervision only. In this work, we\nfurther reduce the labeling effort by allowing such 3D reconstruction methods\nleverage unlabeled images. In order to obtain the viewpoints for these\nunlabeled images, we propose to use a Siamese network that takes two images as\ninput and outputs whether they correspond to the same viewpoint. During\ntraining, we minimize the cross entropy loss to maximize the probability of\npredicting whether a pair of images belong to the same viewpoint or not. To get\nthe viewpoint of a new image, we compare it against different viewpoints\nobtained from the training samples and select the viewpoint with the highest\nmatching probability. We finally label the unlabeled images with the most\nconfident predicted viewpoint and train a deep network that has a\ndifferentiable rasterization layer. Our experiments show that even labeling\nonly two objects yields significant improvement in IoU for ShapeNet when\nleveraging unlabeled examples. Code is available at\nhttps://github.com/IssamLaradji/SSR.\n