.png){kind=icon}
返回
Multi-view and spatial-correlation interaction for multi-scale object detection
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-025-02176-8.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Recent advancements in deep learning have significantly improved object detection performance, yet scale variation remains a formidable challenge. Classical methods, such as Feature Pyramid Network (FPN), often suffer from semantic dilution and inadequate spatial information transmission. To tackle these issues, we propose a novel framework called Multi-view and Spatial- correlation Interaction (MSI) incorporating three key modules: Multi-head Mixed Fusion (MMF), Spatial-Correlation Propagation (SCP), and Scale-Aware Aggregation (SAA). The MMF module enhances high-level semantic features through deep exploration and a refined learning process. The SCP module utilizes high-resolution features to ensure effective spatial information acquisition across different hierarchical levels, thereby improving the spatial positional accuracy of multi-level features. The SAA module overcomes the constraints of traditional pyramid architectures by facilitating both local and global feature aggregation, thereby enhancing the multi-scale representation capability. Extensive experiments on the COCO dataset demonstrate the effectiveness, superiority, and general applicability of our approach. The proposed method significantly improves the utilization of multi-level features and the accuracy of object localization tasks, providing a robust solution to the scale variation problem in object detection.
Keyword:
Object detection
Instance segmentation
Scale variation
Cross-level interaction
期刊
IF:
3.1
论文数:
2.8K
被引数:
2.7K
机构
引用论文
.png){kind=icon}
分享收藏Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive Training Sample Selection通过自适应训练样本选择弥合基于锚点和无锚检测之间的差距
OAAI
OAAI分享收藏
查看更多内容