Reproduce “YOLO11-4K: An Efficient Architecture for Real-Time Small Object Detection in 4K Panoramic Images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Huma Hafeez, Matthew Garratt, Jo Plested, Sankaran Iyer, Arcot Sowmya
Researchers at UNSW developed YOLO11-4K, an efficient architecture for real-time small object detection in 4K panoramic images. It achieves a 75% latency reduction compared to the YOLO11 baseline, processing 4K images at 35.3 FPS, and improves mAP@50 to 0.95 by detecting objects as small as 2.7 x 25.1 pixels.
Key insights
- Directly processing native 4K resolution is critical for small object detection in panoramic images, avoiding detail loss from downsampling or cropping.
- A hybrid backbone strategy (GhostConv in shallow layers, standard Conv in deep layers) optimizes both computational efficiency and GPU execution speed, outperforming full GhostConv replacement.
- A dedicated high-resolution P2 detection layer significantly enhances the model's ability to capture fine spatial details, crucial for detecting very small and distant objects in expansive scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.