Table 1.
Node and hyperedge attribute definitions.
Fig 1.
Hyper-GNN message passing flow chart.
Fig 2.
Feature map resolution enhancement and detail restoration.
Fig 3.
Schematic diagram of dynamic inference path selection.
Fig 4.
Examples of multimodal samples in the OWOD dataset.
Table 2.
Comparison of Hyper-MDR with recent state-of-the-art methods on the OWOD dataset.
Table 3.
Impact of each module on overall performance.
Table 4.
Statistics of the meta-policy controller’s output actions in different scenarios.
Fig 5.
Policy adjustment latency distribution in different scenarios.
Fig 6.
T-SNE visualization of cross-modal feature alignment.
Fig 7.
Time series of attention entropy in modal fusion.
Fig 8.
ROC curve for unknown class detection.
Fig 9.
Pseudo-labeling quality and continuous learning’s ability to resist forgetting.
Fig 10.
Detection results for different object types.
Fig 11.
GFLOPs statistics and efficiency gains of dynamic inference.