Unbiased Heterogeneous Scene Graph Generation with Relation-Aware Message Passing Neural Network

Authors: Kanghoon Yoon, Kibum Kim, Jinyoung Moon, Chanyoung Park

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive evaluations demonstrate that Het SGG outperforms state-of-the-art methods, especially outperforming on tail predicate classes.
Researcher Affiliation Academia 1Dept. of Industrial and Systems Engineering, KAIST, Daejeon, Republic of Korea 2Graduate School of Artificial Intelligence, KAIST, Daejeon, Republic of Korea 3Electronics and Telecommunications Research Institute, 218 Gajeong-ro, Yuseong-gu, Daejeon, Republic of Korea 4 ETRI School, University of Science and Technology, 218 Gajeong-ro, Yuseong-gu, Daejeon, Republic of Korea
Pseudocode No The paper describes the proposed method using prose and mathematical equations, but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes The source code for Het SGG is available at https://github. com/Kanghoon Yoon/hetsgg-torch.
Open Datasets Yes We evaluate Het SGG compared with state-of-the-arts methods on commonly used benchmark datasets, i.e., Visual Genome (VG) (Krishna et al. 2017), and Open Images (OI) V6 (Kuznetsova et al. 2020b).
Dataset Splits Yes A total of 108k images are split into training set (70%) and test set (30%)." and "OI V6 has 301 object classes, and 31 predicate classes, and is split into 126,368 train images, 1,813 validation images, and 6,322 test images.
Hardware Specification No The paper mentions the use of specific models like Res Ne Xt-101-FPN and Faster R-CNN, but does not provide details on the hardware (e.g., specific GPU or CPU models) used for experiments.
Software Dependencies No The paper does not specify version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes For RMP of Het SGG, we set the number of bases (i.e., b) to 8 in VG and 4 in OI, and use four MPNN layers (i.e., l = 4). For SGGen task, we select the top 80 object proposals sorted by object scores, and use per-class non-maximal suppression (NMS) (Zellers et al. 2018) at Io U 0.5.