💻 computer science

Resource-Aware Small-UAV Perception under Annotation and Compute Constraints

This paper proposes and evaluates a resource-aware perception strategy for small UAV detection that combines confidence-guided model routing and difficulty-ranked replay to effectively optimize performance under simultaneous constraints on inference-time computation and annotation budgets.

Original authors: Xufeng Zhou, Jun Li, Hua Hu, Yan Liu, Xiaodong Wang, Kebin Chen, Chao Jiang, Quantong Lv, Xuping Jiang

Published 2026-09-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xufeng Zhou, Jun Li, Hua Hu, Yan Liu, Xiaodong Wang, Kebin Chen, Chao Jiang, Quantong Lv, Xuping Jiang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quiet spaces between the ground and the clouds, a new kind of surveillance is taking shape. Small, unmanned aircraft are becoming common tools for everything from inspecting power lines to monitoring wildlife, but they are also a challenge for safety and security. To keep an eye on them, we need cameras that can spot a tiny, distant speck against a vast, shifting sky. This is a difficult task for a computer. A drone far away might occupy only a few pixels on a screen, easily lost in the visual noise of clouds, buildings, or tree branches. Furthermore, the computers that run these cameras are often small and battery-powered, meaning they cannot afford to run heavy, complex software on every single image they capture. At the same time, teaching these computers to recognize drones usually requires thousands of labeled photos, a resource that is often scarce when a new monitoring mission begins. The core question is how to build a system that sees clearly without needing too much power or too many examples to learn from.

Researchers have tackled this problem by treating the limits of power and data not as obstacles to be overcome, but as conditions to be managed. They tested a strategy on a standard set of drone images, using a popular type of detection software that comes in different sizes, from very light and fast to heavier and more detailed. Their first discovery was straightforward: the size of the image matters more than the size of the software. When they fed the computer images at a higher resolution, allowing it to see more detail, the system became significantly better at finding the drones, even when using a smaller, faster program. The best single setup they found used a medium-sized program running on high-resolution images, which caught nearly 91 percent of the drones in their test set. However, running a heavy program on every single frame is too slow for a small, battery-powered device.

To solve the speed problem, the team introduced a two-step process that acts like a filter. First, a very lightweight, fast program scans every image. If it is confident it sees nothing or sees something clearly, it moves on. But if the fast program is unsure, it passes that specific image to a larger, more powerful program for a second look. This approach, which they tested by routing about 36 percent of the images to the stronger program, improved the system's ability to find drones without slowing it down to the speed of the heavy program alone. The system became much better at catching the drones it had previously missed, raising its overall success rate, while still running at a speed of about 16 frames per second. Crucially, the researchers proved that this improvement came from the smart choice of which images to re-examine, not just from looking at more images. When they randomly sent images to the stronger program instead of using the confidence-based filter, the system performed worse.

The study also addressed the problem of limited training data. In many real-world scenarios, a new monitoring task might start with only half the photos needed for a full training session. The researchers tested a method to make the most of these scarce labels by focusing the computer's learning on the images it found most difficult. They ranked the training photos based on how many drones the computer missed, how poorly it located them, or how small the drones were. By replaying these difficult examples more often during training, the system learned slightly better than if it had just repeated all the photos equally. This benefit was most noticeable when the system was trying to find the tiniest drones, improving its ability to spot them by a small but meaningful margin. However, this trick worked best when the computer already had a decent foundation of knowledge; when the training data was extremely scarce, the system was too confused to know which examples were truly difficult.

The researchers also checked how well their system held up against bad weather and image distortions, such as blur or noise. The system remained remarkably stable, retaining over 96 percent of its accuracy even when the images were corrupted. They even tested whether a large, advanced artificial intelligence model could help by looking at the whole picture to find drones, but found that these models were unreliable for this specific task without the initial help of the camera-based detector. The final picture that emerges is one of balance. The most effective approach for spotting small drones on limited hardware is not to use the biggest possible model, but to use a smart combination of a fast first pass and a careful second look, supported by high-resolution images and a training process that focuses on the hardest examples. This resource-aware strategy allows a small computer to see what it needs to see, proving that careful management of limited resources can be just as powerful as raw computing power.

Technical Summary: Resource-Aware Small-UAV Perception under Annotation and Compute Constraints

Problem Statement
Small unmanned aerial vehicle (UAV) detection faces a dual resource constraint: limited computational power during inference (often on edge devices or CPUs) and limited availability of annotated training data (domain-specific cold starts). While existing object detection benchmarks and model families (e.g., YOLO) offer robust architectures, they do not inherently address the specific trade-offs required for tiny UAVs, which occupy few pixels and are susceptible to background clutter. Furthermore, standard evaluations often ignore the interplay between inference-time computation and the efficiency of learning from sparse annotations. This study addresses how to allocate these limited resources—computation and annotation—more deliberately to improve detection performance without relying on additional UAV-domain pretraining or external data.

Methodology
The study employs a controlled evaluation framework using the visible-light detection subset of the DUT-Anti-UAV benchmark. The methodology is divided into three primary components:

Baseline Establishment: The authors first established baselines by comparing detector scales (YOLOv8n, s, m; YOLOv9; YOLO11) and input resolutions (320 to 640) under a "domain-specific cold start" protocol. This protocol initializes models with generic public pretrained weights and fine-tunes them solely on the DUT-Anti-UAV split, without additional UAV-specific pretraining.
Confidence-Guided Adaptive Inference: To address compute constraints, the study implements a two-stage cascade. A lightweight detector (YOLOv8n) processes all images. Images with low confidence scores ( $q_j < \eta$ ) are routed to a larger verifier (YOLOv8m). This "confidence-guided routing" is rigorously compared against a "random routing" control, where the same fraction of images is routed to the larger model without using confidence scores, to isolate the benefit of sample selection from the benefit of increased compute volume.
Difficulty-Ranked Hard-Example Replay: To address annotation constraints, the study evaluates training strategies under reduced label budgets (10%, 25%, 50%). After initial supervised training, a "difficulty-ranked replay" strategy is applied. Images are scored based on missed objects, localization error, low matched confidence, tiny-target exposure, and false-positive pressure. These scores determine the multiplicity of replay (1x, 2x, or 3x copies) during a short additional training phase. This is compared against "plain continuation" (more training on the same data) and "compute-matched uniform replay" (randomly replaying data with the same total sample count).

Key Contributions
The paper makes three specific contributions:

Confidence-Guided Adaptive Computation: It provides a method for routing uncertain images to heavier models, validated against a matched random-routing control to prove that gains stem from intelligent sample selection rather than mere increased compute.
Difficulty-Ranked Hard-Example Replay: It evaluates a heuristic difficulty ranking for replaying labeled data, demonstrating its specific utility in improving tiny-target recall under moderate label budgets (50%), while distinguishing this effect from general gains due to increased training volume.
Integrated Resource-Aware Evaluation: It supplies a framework that jointly considers detector scale, input resolution, annotation budget, target size, and statistical uncertainty, rather than treating these as isolated variables.

Results

Detector and Resolution: Input resolution had a more significant impact on performance than changing detector families at low resolutions. At 640 input size, YOLOv8s achieved the highest observed mAP50 of 0.9052.
Adaptive Inference: At a 416-input resolution with a 35.68% routing rate, confidence-guided routing increased the F1 score from 0.8132 (YOLOv8n alone) to 0.8446. This outperformed random routing at the same rate (F1 0.8237), confirming the value of confidence-based selection. The analytical cascade throughput was estimated at 16.00 FPS on a CPU.
Limited Annotation: Under a 50% label setting, difficulty-ranked replay provided a moderate improvement over compute-matched uniform replay, increasing mAP50 from 0.7575 to 0.7621 and tiny-target recall from 0.5730 to 0.5861. This benefit was not stable at lower label budgets (10% or 25%), suggesting the strategy requires a sufficiently stable initial representation to be effective.
Robustness and Uncertainty: Model scaling (n to s to m) provided the largest gains for tiny and small targets. The models retained approximately 96.5% of their clean-image mAP50 under synthetic corruptions, with Gaussian noise and overexposure being the most challenging conditions. Bootstrap analysis indicated that while larger models performed better, the differences were often within statistical uncertainty ranges, supporting a view of detector choice as a practical trade-off.

Significance and Claims
The paper claims that resource constraints should be treated as an intrinsic part of the perception problem rather than external limitations. The results support a detector-first resource allocation strategy:

Resolution and Capacity: Input resolution and model capacity must be selected jointly with the target-size distribution.
Inference Allocation: Confidence-guided routing effectively allocates inference compute, recovering recall on difficult frames without the cost of running heavy models on every frame.
Training Allocation: Difficulty-ranked replay offers a targeted way to utilize limited annotations, specifically improving recall for tiny targets once the detector has learned a stable representation.

The authors explicitly note that the throughput figures are analytical estimates based on standalone CPU timings and that the VLM (Vision-Language Model) branch was used only as a diagnostic for applicability boundaries, not as a primary performance driver. The study concludes that its findings are bounded by the single dataset used, the recorded-image evaluation nature, and the lack of end-to-end hardware deployment measurements.

More like this