Scene Graph Detection on Visual Genome (VG)
36.5R@50FlowSG
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| FlowSGMethod Stages=Two-stage, Evaluation Protocol=closed-set2026.04 | 36.5 | 42.4 | 18.4 | 21.6 | — | — | |
| SS R-CNNBackbone=ResNeXt-101-FPN2023.03 | 33.5 | 38.4 | 8.6 | 10.3 | 21.1 | 24.4 | |
| RU-NetBackbone=ResNeXt-101-FPN2023.03 | 32.9 | 37.5 | — | 10.8 | — | 24.2 | |
| DSGGMethod Stages=Two-stage, Evaluation Protocol=closed-set2026.04 | 32.9 | 38.5 | 13 | 17.3 | — | — | |
| OpenPSGMethod Stages=Two-stage, Evaluation Protocol=closed-set, reproduced=true2026.04 | 32.7 | 38 | 13.5 | 18.3 | — | — | |
| MOTIFMethod Stages=Two-stage, Evaluation Protocol=closed-set2026.04 | 32.5 | 36.8 | 6.6 | 7.9 | — | — | |
| PE-Net(P)Backbone=ResNeXt-101-FPN, Training Protocol=PL only2023.03 | 32.4 | 36.9 | 8.9 | 11 | 20.7 | 24 | |
| MotifsBackbone=ResNeXt-101-FPN2023.03 | 32.1 | 36.8 | 6.6 | 7.9 | 19.4 | 22.4 | |
| VCTreeBackbone=ResNeXt-101-FPN2023.03 | 31.9 | 36 | 6.4 | 7.3 | 19.2 | 21.7 | |
| VCTreeMethod Stages=Two-stage, Evaluation Protocol=closed-set2026.04 | 31.9 | 36 | 6.4 | 7.3 | — | — | |
| GPS-NetBackbone=ResNeXt-101-FPN, Reproduction=codebase [13], [27]2023.03 | 31.3 | 35.9 | 6.7 | 8.6 | 19 | 22.3 | |
| BGNNBackbone=ResNeXt-101-FPN, Reproduction=codebase [13], [27]2023.03 | 31 | 35.8 | 10.7 | 12.6 | 20.9 | 24.2 | |
| PE-NetBackbone=ResNeXt-101-FPN, Training Protocol=PL and PR2023.03 | 30.7 | 35.2 | 12.4 | 14.5 | 21.6 | 24.9 | |
| GR-CNNBackbone=ResNeXt-101-FPN, Reproduction=codebase [13], [27]2023.03 | 29.7 | 32.8 | 5.8 | 6.6 | 17.8 | 19.7 | |
| VTransEBackbone=ResNeXt-101-FPN2023.03 | 29.7 | 34.3 | 5 | 6.1 | 17.4 | 20.2 | |
| R-CAGCNBackbone=ResNeXt-101-FPN2023.03 | 28.1 | 31.3 | 7.9 | 8.8 | 18 | 20.1 | |
| Motifs-NICEBackbone=ResNeXt-101-FPN2023.03 | 27.8 | 31.8 | 12.2 | 14.4 | 20 | 23.1 | |
| KERNBackbone=ResNeXt-101-FPN, Reproduction=codebase [13], [27]2023.03 | 27.1 | 29.8 | 6.4 | 7.3 | 16.8 | 18.6 | |
| PE-Net-ReweightBackbone=ResNeXt-101-FPN, Training Protocol=PL and PR, Re-weighting Strategy=[2]2023.03 | 26.5 | 30.9 | 16.7 | 18.8 | 21.6 | 24.9 | |
| PE-NetMethod Stages=Two-stage, Evaluation Protocol=closed-set2026.04 | 26.5 | 30.9 | 16.7 | 18.8 | — | — | |
| Motifs-ReweightBackbone=ResNeXt-101-FPN, Re-weighting Strategy=[2]2023.03 | 24.9 | 26.3 | 13.3 | 15.4 | 19.2 | 21.8 | |
| SGTRBackbone=ResNeXt-101-FPN2023.03 | 24.6 | 28.4 | 12 | 15.2 | 18.3 | 21.8 | |
| Motifs-BPL-SABackbone=ResNeXt-101-FPN2023.03 | 23 | 26.9 | 13.5 | 15.6 | 18.3 | 21.3 | |
| Motifs-PPDLBackbone=ResNeXt-101-FPN2023.03 | 21.2 | 23.9 | 11.4 | 13.5 | 16.3 | 18.7 | |
| Motifs-CogTreeBackbone=ResNeXt-101-FPN2023.03 | 20 | 22.1 | 10.4 | 11.8 | 15.2 | 17 | |
| Motifs-GCLBackbone=ResNeXt-101-FPN2023.03 | 18.4 | 22 | 16.8 | 19.3 | 17.6 | 20.7 | |
| CAPSGGMethod Stages=Two-stage, Evaluation Protocol=closed-set2026.04 | 18.3 | 22.8 | — | — | — | — | |
| Motifs-TDEBackbone=ResNeXt-101-FPN, Strategy=TDE2023.03 | 16.9 | 20.3 | 8.2 | 9.8 | 12.6 | 15.1 |