Optimization of object detection models for low-altitude scenarios
Abstract
Target detection in low-altitude scenes is challenging due to low recognition rates for small targets and feature extraction difficulties. We propose the DCSOD-YOLO model, incorporating dynamic convolution and a novel small object detection head, trained on a specific low-altitude dataset. On the VisDrone and our low-altitude datasets, our model achieved 48.2 mAP with 52.3M parameters and 22.7 ms@640² (Nvidia 4070), outperforming baselines by reducing parameters while enhancing accuracy and speed.
Fields of Study & Subjects
Advanced Neural Network Applications
Domain Adaptation and Few-Shot Learning
Visual Attention and Saliency Detection
Publication details
- Journal / source
- International Conference on Graphics and Image Processing (ICGIP)
- Pages
- 105–105
- Publication date
- 4 March 2026
- Language
- EN