Multi-scenario object detection model evaluation method based on hierarchical game theory

To address the limitations of existing evaluation methods for object detection models, which lack a structured evaluation framework and weight modelling mechanism, this paper proposes a comprehensive multi-level, multi-scenario evaluation approach. The method constructs a two-tier evaluation metric system, including basic performance, efficiency, robustness, and security. For weight modelling, the approach integrates Shapley value theory and expert knowledge to quantify the contribution of each metric, and employs a Bayesian updating strategy based on Dirichlet distribution to enable dynamic weight modelling. Evaluation experiments were conducted on seven representative models using the VOC2007 dataset. The results show that the proposed method achieves a Kendall and Spearman rank correlation coefficient of 1.00 in typical scenarios, and 0.90 and 0.96 in boundary scenarios, respectively, outperforming other comparison methods. This validates the effectiveness of the proposed approach.