HyMTAS-GBM: Integrating Semantic Convolutions and Multi-Head Attention with Gradient Boosting for Robust Object Detection

Abstract

Object detection in complex scenes requires models that can maintain precise spatial attention and reliable query representations. However, existing CNN and DETR-based detectors often exhibit diffuse cross-attention, weak query initialization, and limited robustness in cluttered or dynamic environments. This paper proposes HyMTAS-GBM, a hybrid multi-head transformer architecture that integrates attention-consensus learning, semantic-guided query selection, and gradient-boosting classification to improve detection reliability across image and video domains. The framework introduces Consensus Multi-Head Attention (CoMHA) to stabilize attention correlations, Semantic-Regional Query Selection (SRQS) for object-aware query initialization, and Regional Reinforcement Cross-Attention (RRCA) with learnable key-point agents for spatially constrained reasoning. A sparsity-guided feature amplification pipeline further enhances encoder representations for small and occluded objects. Extensive experiments on ScanNetV2, SUN RGB-D, ImageNet VID, and UA-DETRAC demonstrate consistent improvements, achieving 90.6 mAP25 and 85.3 mAP50 on ScanNetV2 while maintaining competitive efficiency.