Learning Canonical Representations for Scene Graph to Image Generation

2020-12-30 2020-12-30 约 129 字预计阅读 1 分钟

Learning canonical representations for scene graph to image generation

ECCV2020 European conference on computer vision

Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, Amir Globerson

Tel Aviv University, UCB, Bar-Ilan University NVIDIA

semantic equivalence

Motivation

过去的sg2im的一个不足是不能捕捉graphs中的语义等价性(semantic equivalence)
- 即：同样一张图片可以用多个逻辑上等价的SG来表述
所以提出从数据中学习出canonical graph representations
主要展示3个数据集：visual genome, COCO, clevr