Learning Direct Optimization for scene understanding
作者：
Highlights：
• We explain a single image with an interpretable 3D computer graphics scene model.
• We allow for the joint refinement of all the latent variables in a multi-object scene.
• Our method (LiDO) does not require a specific error metric to compare images.
• LiDO is faster and generally is more stable than standard optimizers.
• LiDO deals well with a mismatch between the real images and the fitted scene model.
摘要
•We explain a single image with an interpretable 3D computer graphics scene model.•We allow for the joint refinement of all the latent variables in a multi-object scene.•Our method (LiDO) does not require a specific error metric to compare images.•LiDO is faster and generally is more stable than standard optimizers.•LiDO deals well with a mismatch between the real images and the fitted scene model.
论文关键词：Computer vision,Scene understanding,3D Reconstruction,Inverse graphics,Object recognition,Scene graph,Analysis-by-synthesis,Graphics
论文评审过程：Received 10 June 2019, Revised 29 January 2020, Accepted 6 April 2020, Available online 23 April 2020, Version of Record 3 May 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107369
