Od-Dda: Real-Time Object Detector with Dual Dynamic Adaptation in Variable Scenes

Abstract

Object detection in complex scenarios faces numerous challenges, including dense small objects, adverse weather conditions, and varied background interference. We propose an Object Detector with Dual Dynamic Adaptation (OD-DDA) to address these challenges, enabling the network to adapt to unpredictable feature changes rapidly. First, at each stage of the backbone network, the Dual Dynamic Adaptation Network (DDANet), we introduce the Dynamic Feature Adaptation (DFA) module, which quickly captures multi-scale local contextual information to respond to unpredictable feature changes. In addition, to utilize the complex background information more efficiently, we design the Dynamic Fine-Grained Weight Adaptation (DFGWA) module. The DFGWA selectively learns weights based on a fine-grained feature hierarchy of contextual information, which improves the response speed to small objects. In addition, we also introduce the Efficient Weighted Adaptive Downsampling (EWAD) technique in this module to reduce redundant information and speed up network inference. Through the synergy of these modules, OD-DDA can handle the detection problem of complex scenes more flexibly and significantly improve the inference speed. We conduct rigorous experimental comparisons on five datasets, and the results show that OD-DDA performs excellently in various scenarios. Especially on the UAVDT dataset, AP50 reaches 37.9% and FPS reaches 87.5, proving its ability to balance speed and accuracy.

Keywords: Object Detection, Dynamic adaptation, Fine-grained features, Efficient network, Contextual information