Table 4 Performance comparison (%) of our method with state-of-the-art methods on the GQA dataset
Models | Predicate classification | Scene graph classification | Scene graph detection | ||||||
|---|---|---|---|---|---|---|---|---|---|
mR@50/100 | R@50/100 | F@50/100 | mR@50/100 | R@50/100 | F@50/100 | mR@50/100 | R@50/100 | F@50/100 | |
MOTIFS (Zellers et al. 2018) | 16.4 / 17.1 | 65.3 / 66.8 | 26.2 / 27.2 | 8.2 / 8.6 | 33.4 / 34.2 | 13.0 / 13.9 | 5.8 / 6.6 | 27.2 / 30.7 | 9.6 / 10.9 |
MOTIFS+DC (Han et al. 2022) | 21.4 / 22.5 | 61.3 / 62.7 | 31.7 / 33.1 | 9.9 / 10.4 | 32.5 / 33.1 | 15.2 / 15.8 | 9.4 / 10.7 | 28.3 / 32.1 | 14.1 / 16.1 |
MOTIFS+CFA (Li et al. 2023a) | 31.7 / 33.8 | – / – | – / – | 14.2 / 15.2 | – / – | – / – | 11.6 / 13.2 | – / – | – / – |
MOTIFS+EICR (Min et al. 2023) | 36.3 / 38.0 | 56.4 / 58.1 | 44.2 / 46.0 | 17.2 / 18.2 | 28.8 / 29.4 | 21.5 / 22.5 | 16.0 / 18.0 | 24.6 / 28.4 | 19.4 / 22.0 |
MOTIFS+AMCA(Ours) | 39.8 / 41.0 | 47.0 / 48.9 | 43.1 / 44.6 | 18.5 / 19.7 | 27.8 / 28.7 | 22.2 / 23.4 | 17.7 / 19.8 | 20.4 / 24.0 | 19.0 / 21.7 |
VCTree (Tang et al. 2019) | 16.6 / 17.1 | 63.8 / 65.7 | 26.3 / 27.5 | 7.9 / 8.3 | 34.1 / 34.8 | 12.8 / 13.4 | 6.5 / 7.4 | 28.3 / 31.9 | 10.6 / 12.0 |
VCTree+DC (Han et al. 2022) | 22.2 / 23.5 | 61.3 / 62.9 | 32.6 / 34.2 | 10.5 / 11.2 | 32.2 / 32.8 | 15.8 / 16.7 | 8.0 / 9.5 | 25.8 / 29.1 | 12.2 / 14.3 |
VCTree+CFA (Li et al. 2023a) | 33.4 / 35.1 | – / – | – / – | 14.1 / 15.0 | – / – | – / – | 10.8 / 12.6 | – / – | – / – |
VCTree+EICR (Min et al. 2023) | 35.9 / 37.4 | 55.3 / 57.0 | 43.5 / 45.2 | 17.8 / 18.6 | 28.4 / 29.1 | 21.9 / 22.7 | 14.7 / 16.3 | 24.0 / 27.6 | 18.2 / 20.5 |
VCTree+AMCA(Ours) | 37.8 / 40.3 | 45.1 / 47.0 | 41.1 / 43.4 | 21.0 / 21.9 | 29.4 / 30.3 | 24.5 / 25.4 | 16.4 / 18.6 | 19.0 / 22.5 | 17.6 / 20.4 |