NGP
Scene Graph Generation Using Background Knolwedge
NGP is a technique implemented in PyTorch for regularizing deep models for scene graph generation (SGG) at training-time, by injecting commonsense knowledge. NGP achieves the following: (i) it improves the accuracy of IMP, MOTIFS, and VCTree baseline SGG models, by up to 33%, leading to 16% absolute accuracy improvements when applied in conjunction with TDE; (ii) it outperforms GLAT and LENSR, two state-of-the-art training-time regularization techniques, by up to 18% and 15%; (iii) it can improve the accuracy of a baseline SGG model by up to six times when restricting the availability of ground-truth training facts; (iv) it outperforms in accuracy BGNN by up to 90% and KBFN by up to 86% when combined with TDE. NGP has been evaluated on the Visual Genome and the Open Images v6 benchmarks. This library builds upon Mask R-CNN for bounding box detection.
Repository
Please contact me to grant you access.