Object detection ensembles often rely on post-processing methods such as Weighted Boxes Fusion (WBF) to combine overlapping predictions from multiple detectors. While effective, standard WBF assumes fixed intersection-over-union (IoU) thresholds and uniform model trust, limiting its adaptability to diverse object scales and varying detector quality. In this work, we introduce Adaptive Reliability-Weighted Boxes Fusion (AR-WBF), a context-aware extension of WBF that addresses these limitations through two key mechanisms. First, each model’s predictions are scaled by a reliability factor reflecting its validated detection accuracy, enabling more trustworthy models to exert greater influence during fusion. Second, AR-WBF employs an adaptive IoU threshold and deferred reliability weighting, calibrating final confidence after spatial aggregation. This strategy preserves geometric consensus while improving recall and confidence calibration. Experiments on the COCO dataset demonstrate that AR-WBF maintains precision and improves recall stability compared to baseline WBF, particularly in heterogeneous ensembles. Moreover, we interpret the adaptive weighting process through a causal reasoning lens, viewing reliability and contextual adaptation as influencing factors in fusion outcomes. AR-WBF thus connects classical statistical fusion with causally inspired, context-aware object detection. Beyond performance gains, it illustrates how lightweight causal reasoning—via reliability priors and contextual interventions—can enhance interpretability and robustness in ensemble perception systems.
Daoud et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: