Multi-Scale Feature Aligned for Object Detection

Abstract

Existing object detection methods face challenges in multi-scale detection, especially in simulation environments with large object size variability. First, feature differences across scales hinder matching and affect accuracy. Second, inconsistent features prevent full use of multi-scale information, reducing performance. To address this, the authors propose the multi-scale feature aligned detection model. To enhance matching, the authors introduce a multi-scale query matching mechanism, using learnable queries and cross-attention to dynamically select optimal scales, improving robustness. To tackle regression-classification inconsistency, the authors propose a feature-category alignment strategy, leveraging the contrastive language-image pre-training model to align object and category features and reduce classification mismatches. Experiments show multi-scale feature aligned detection model outperforms mainstream methods with notable gains in multi-scale and complex simulation scenarios. This approach offers a robust, efficient solution for simulation-driven detection tasks needing high precision across varied object scales.