Introduction
Detecting drones in high-definition imagery is a significant challenge, often described as a "needle in a haystack" problem. A distant drone may occupy less than 1% of the total pixels in a standard image. Traditional computer vision and deep learning models often struggle with these "extremely small" objects. As neural networks deepen to extract high-level semantic features, the fine-grained spatial details—such as the thin lines of a drone's frame—frequently vanish.
Furthermore, drone detection is increasingly required at the "edge," necessitating deployment on portable devices or other drones with limited memory and processing power. Heavyweight models like YOLOv3 or Faster R-CNN are often too memory-intensive or fail to capture the minute details of a 12x7 pixel object against complex backgrounds like trees or buildings.
Methodology: The TIB-Net Architecture
TIB-Net is an ultra-lightweight detection network built upon the Extremely Tiny Face Detector (EXTD). It utilizes an iterative backbone to maintain a low parameter count while introducing two primary innovations for drone detection:
- Cyclic Pathway: This structure transmits low-level features (textures and contours) from earlier iterations directly into higher-level layers. By "recycling" these sharp details, the network ensures it does not "forget" the characteristics of a tiny drone as it processes more abstract information.
- Spatial Attention Module (SAM): Because drones lack internal texture, spatial context is the most critical signal for detection. The SAM helps the network focus on informative spatial regions and suppress irrelevant background noise, such as birds or distant windows.
- Basic Component Block (BCB): These blocks integrate inverted residuals with the spatial attention module, maintaining efficiency without sacrificing the ability to focus on small targets.
A key finding in the development of TIB-Net was the "No Updating" strategy. Ablation studies revealed that keeping the original low-level feature map unchanged and feeding it into every cyclic pathway outperformed iterative updates. This suggests that for tiny objects, preserving raw structural information is more beneficial than allowing it to be modified by the network's learning process.
Dataset and Analysis
Due to the scarcity of public drone datasets, the authors developed the Drone Benchmark Dataset. This collection includes 2,860 images of multi-rotor and fixed-wing drones captured under various lighting conditions (day, nightfall, night) and against complex backgrounds. Many drones were captured at distances of approximately 500 meters, resulting in extremely small and often blurred targets.
As shown in the distribution of drone sizes, a significant portion of the dataset consists of objects with a bounding box area of less than 400 pixels, highlighting the difficulty of the detection task.
Key Results
TIB-Net demonstrates that efficiency does not have to come at the cost of accuracy. The model achieved a state-of-the-art 89.2% mean Average Precision (mAP), outperforming the base EXTD (85.1%) and YOLOv3 (84.9%) on the custom dataset.
The most striking feature of TIB-Net is its size. The final model occupies only 697.0 KB, making it small enough to fit on almost any embedded system. In comparison, YOLOv3 is roughly 234 MB—over 300 times larger. Qualitative results further confirm the model's robustness, successfully detecting drones that are heavily blurred or blended into dark, night-time backgrounds.
Limitations and Future Work
While TIB-Net is exceptionally efficient in terms of memory, its inference speed is currently 0.290s per image (~3.4 FPS) on a high-end Titan Xp GPU. This does not yet meet the requirements for high-speed, real-time tracking, which typically requires 30+ FPS. Future research will focus on speed optimization through pruning and architectural tweaks, as well as improving the network's ability to distinguish between drones and other small moving objects like birds in chaotic environments.
Summary updated