Abstract
Detecting tiny objects in satellite imagery, such as vehicles or small vessels, remains a significant challenge for modern deep learning architectures. Standard Feature Pyramid Networks (FPNs) often suffer from "interfering activations"—background noise in high-level branches that mimics large objects and overwhelms the signal of small targets. This paper introduces Renormalized Connections (RC), a lightweight, physics-inspired plug-in designed to harmonize multi-scale features. By applying principles from renormalization group theory, RC enables features to work synergistically rather than independently. Experimental results across five diverse datasets demonstrate that RC significantly boosts detection accuracy for tiny objects with virtually zero additional computational cost.
The Challenge of Scale-Preferred Detection
In remote sensing, objects of interest are frequently just a few pixels in size. Traditional object detectors employ a "divide-and-conquer" strategy via FPNs, where different network layers are assigned to detect objects of different scales. However, for "scale-preferred" tasks—where the objective is exclusively tiny objects—this strategy falters.
The authors observed that high-level FPN branches, intended for large objects, generate massive amounts of background noise (e.g., from roads or buildings) when no large objects are present. This noise interferes with the training of small-object branches, a problem that standard loss functions like Focal Loss cannot adequately resolve.
Methodology: Renormalized Connections
To address this, the researchers proposed a two-step solution inspired by Renormalization Group Theory, which studies physical systems at varying scales.
1. Knowledge Discovery Network (KDN)
The team first developed a single-branch "probe" network to analyze the relationships between features at different backbone stages. They discovered that while these stages are structurally independent, they can be combined to act synergistically to suppress noise and enhance target signals.
2. Renormalized Connection (RC) Architecture
Based on KDN findings, the authors abstracted a linear connection logic termed n21C. The most effective variant, 421C (or E421C), uses a specific 4:2:1 ratio to combine features from different levels. Unlike simple feature addition, RC constructs "independent bases" by subtracting correlated parts of different layers. This ensures the information reaching the detection head is "renormalized"—cleared of redundant background noise and focused on the target scale.
Experimental Results and Insights
The effectiveness of RC was validated across five datasets, including the new IPIU Dataset (focused on tiny moving vehicles) and standard benchmarks like RSOD and NWPU VHR-10.
Noise Suppression
Visual analysis using heatmaps confirms that the standard PAFPN architecture produces significant interfering activations in its high-level branches (P4 and P5). When the 421C connection is integrated, these false activations are drastically reduced, allowing the model to focus exclusively on the actual targets.
Performance Gains
- Accuracy: On the challenging IPIU task, adding 421C to Faster R-CNN improved mAP from 13.4% to 26.3%. For YOLOv8, mAP rose from 55.1% to 55.5%.
- Efficiency: The linear E421C version adds almost no parameters. YOLOv8 with RC maintained a real-time inference speed of 16.89 ms on a GTX 1080 GPU.
- Versatility: The method proved robust across 17 different architectures, including FCOS, RetinaNet, and Swin Transformers.
Conclusion
Renormalized Connections offer a simple yet powerful "Occam’s Razor" solution for satellite object detection. By reorienting gradient flow and filtering redundant noise through physics-inspired linear combinations, RC allows standard detectors to excel at tiny object detection without the need for heavy architectural redesigns. Future work will explore applying these connections to even more complex multi-scale remote sensing missions.
Summary updated