💻 computer science

SPQ-DETR for Tiny-UAV Detection]{SPQ-DETR: Prior-Guided Queries and Shape-Aware Geometric Supervision for Long-Range Tiny-UAV Detection

This paper proposes SPQ-DETR, a transformer-based detector that enhances long-range tiny-UAV detection by integrating high-resolution feature preservation, prior-guided queries initialized from encoder tokens, and shape-aware geometric supervision to significantly improve accuracy and recall on challenging datasets.

Original authors: Guangshuo Zhang, Yunpeng Hu, Ying He, Teng Yu

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guangshuo Zhang, Yunpeng Hu, Ying He, Teng Yu

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

Imagine you are trying to find a tiny, specific speck of dust floating in a massive, chaotic room filled with swirling leaves, moving shadows, and flickering lights. Now, imagine that speck of dust is a drone (a tiny UAV), and the room is a long-distance camera view of a city or the sky.

This is the challenge the paper "SPQ-DETR" tackles. It proposes a new computer vision system designed specifically to spot these microscopic drones that are often lost in the background noise.

Here is how the paper explains their solution, broken down into simple concepts:

The Problem: The "Needle in a Haystack"

Standard security cameras and AI detectors are great at finding big things, like cars or people. But when a drone is far away, it might only take up a few pixels on the screen—like a single grain of sand on a beach.

The Issue: As standard AI looks at an image, it often "zooms out" to understand the big picture. In doing so, it accidentally blurs or loses those tiny grains of sand.
The Result: The AI misses the drone entirely, or it draws a box around it that is wobbly and inaccurate because it can't "feel" the edges of such a small object.

The Solution: SPQ-DETR

The authors built a new system based on a popular AI model called RT-DETR (a fast, smart detector). They added three special "upgrades" to make it better at finding tiny drones.

1. The "High-Resolution Glasses" (DLK-GateNet & P2-augmented Neck)

Imagine trying to read a tiny label on a bottle from across the room. If you squint or look at the whole room, you miss the text.

What they did: They built a special "backbone" (the part of the AI that sees the image) that refuses to throw away the fine details.
The Analogy: Instead of just looking at the whole room, this system keeps a pair of "high-resolution glasses" on the table. It ensures that even the smallest, most detailed parts of the image (the tiny pixels where the drone might be) are preserved and not blurred out by the rest of the processing.

2. The "Smart Search Party" (Prior-Guided Queries)

In the original AI model, the "searchers" (called queries) are like a group of detectives sent into a room to find a suspect. They wander around randomly, hoping to stumble upon the suspect.

The Problem: If the suspect is hiding in a tiny corner, the detectives might miss them because they are looking at the big, obvious things first.
What they did: They gave the detectives a "hint." Before the search starts, the system scans the high-resolution "glasses" to find the most suspicious-looking spots (the places that look most like a tiny drone).
The Analogy: Instead of sending the detectives to wander randomly, the system points them directly to the top 10 most suspicious corners. It says, "Start your search here!" This ensures the AI doesn't waste time looking at empty sky and focuses immediately on the tiny, hard-to-see targets.

3. The "Smooth Ruler" (Shape-Aware Geometric Supervision)

When the AI tries to draw a box around a drone, it uses math to measure how well the box fits. Standard math (like measuring overlap) is very sensitive. If the box moves just one pixel, the score might crash from "perfect" to "terrible," confusing the AI.

The Problem: For a tiny object, a one-pixel shift is a huge mistake, but standard math treats it like a disaster, making the AI unstable.
What they did: They introduced a new way of measuring fit called Shape-NWD.
The Analogy: Imagine trying to measure the distance between two tiny dots. A standard ruler might snap if you move it slightly. The new "Smooth Ruler" is flexible; it tells the AI, "You're close, just move a tiny bit more," rather than screaming "You failed!" This helps the AI learn to draw perfect boxes around tiny objects without getting confused by tiny errors.

The Results

The authors tested this new system on three different datasets (collections of drone videos) that represent different real-world scenarios:

Ground-to-Air: Looking up from the ground at drones in the sky.
Air-to-Air: One drone filming another drone.
Infrared: Seeing drones through heat sensors (where they look like faint blobs).

The Outcome:
Compared to the best existing models, SPQ-DETR found significantly more drones (higher "Recall") and drew tighter, more accurate boxes around them (higher "AP").

It found 2.1 to 2.7% more drones than the previous best models.
It reduced the number of "missed detections," meaning fewer tiny drones slipped through the cracks.

Summary

In short, the paper presents a smarter way to hunt for tiny drones. It keeps the image details sharp (so the drone doesn't disappear), gives the AI a head start by pointing it to the right spots (so it doesn't miss the target), and uses a gentler way of measuring success (so the AI learns to draw perfect boxes). The result is a system that is much better at spotting those tiny, hard-to-find intruders in a cluttered world.

Technical Summary: SPQ-DETR for Long-Range Tiny-UAV Detection

1. Problem Statement

Reliable detection of micro-Unmanned Aerial Vehicles (UAVs) is critical for low-altitude security but remains a significant challenge in long-range surveillance scenarios. In these settings, UAVs often appear as tiny, low-contrast targets occupying only a few pixels, submerged within cluttered and dynamic backgrounds (e.g., clouds, urban textures, foliage).

The paper identifies two fundamental bottlenecks in existing vision-based detectors for this task:

Loss of High-Resolution Cues: Progressive downsampling in standard convolutional or transformer backbones dilutes fine-grained spatial information, causing tiny targets to be lost in deep feature maps. This leads to high rates of missed detections (low Recall).
Unstable Geometric Supervision: Standard overlap-based losses (IoU, GIoU) provide unstable or insufficiently smooth gradients for extremely small bounding boxes. A minor shift in center or scale can cause drastic changes in overlap metrics, degrading optimization stability and localization accuracy.

While multi-modal sensors (radar, RF) exist, they often lack the appearance details required for fine-grained recognition. Vision-based systems are preferred for their richness and low cost but require specialized architectures to handle the "few-pixel" nature of long-range UAVs.

2. Methodology: SPQ-DETR

The authors propose SPQ-DETR (Small-object Prior Query DETR), an end-to-end transformer-based detector built upon the RT-DETR framework. The architecture introduces three coordinated modifications to address the identified bottlenecks:

A. High-Resolution Representation (DLK-GateNet & PCMF Neck)

To preserve fine-grained textures and prevent the dilution of tiny target cues:

DLK-GateNet Backbone: A task-oriented backbone replacing the standard ResNet-50. It utilizes Large-Select Kernel (LSK) inspired blocks with spatial gating. The design employs parallel large-kernel depthwise convolutions (including dilated convolutions) to enlarge the receptive field while using a spatial gate (derived from channel-wise statistics) to suppress background noise and emphasize discriminative tiny-object cues. Crucially, downsampling is minimized within stages to preserve stride-4 details.
P3-Centered Multi-Scale Fusion (PCMF) Neck: A specialized neck that reorganizes backbone features into a pyramid $\{P2, P3, P4, P5\}$ ${P 2, P 3, P 4, P 5}$ .
- It explicitly preserves stride-4 details via a dedicated P2 branch, which is critical for few-pixel UAVs.
- It employs a P3-centered tri-scale fusion hub that injects high-resolution details from $C2$ into the stride-8 space, combining fine spatial cues with mid-level structures and global semantics.
- A MFFF (Multi-scale Feature Fusion and Enhancement) module selectively enhances high-frequency discriminative components while suppressing background entanglement.

B. Prior-Guided Query Initialization (Small-Object Prior Queries)

To guide the decoder toward suspicious tiny-object regions and reduce missed detections:

Instead of relying solely on learnable global queries, SPQ-DETR initializes a subset of decoder queries ( $k_s$ ) from Top-K high-resolution encoder tokens.
Mechanism: The model mines high-confidence tokens specifically from the high-resolution levels ( $P2$ and $P3$ ) based on an objectness score. These tokens are converted into Small-Object Prior Queries with initialized content and reference boxes.
Level Bias: During decoding, the cross-attention for these prior queries can be biased to attend preferentially to high-resolution levels, further focusing the model on tiny candidates.
The remaining queries ( $k_g$ ) remain learnable global embeddings to maintain holistic scene understanding.

C. Shape-Aware Geometric Supervision (Shape-NWD)

To stabilize the optimization of extremely small bounding boxes:

The paper replaces standard IoU-based metrics with Shape-normalized Wasserstein Distance (Shape-NWD) in both the bounding-box regression loss and the Hungarian matching cost.
Rationale: IoU is highly sensitive to tiny perturbations for small boxes. Shape-NWD models boxes as 2D Gaussians, providing a continuous, smooth geometric signal that is less prone to oscillatory gradients.
Consistency: By integrating Shape-NWD into both the assignment criterion (matching) and the regression objective, the paper ensures the optimization goal and the matching strategy are geometrically consistent, stabilizing training for tiny boxes.

3. Key Contributions

SPQ-DETR Framework: An end-to-end transformer detector tailored for long-range tiny-UAV detection, demonstrating effectiveness across diverse surveillance settings (ground-to-air, air-to-air, and infrared).
Small-Object Prior Query Mechanism: A novel decoder-side initialization strategy that leverages high-resolution features to seed a portion of queries with high-confidence candidates. This design consistently boosts Recall and AP50 by reducing missed detections in cluttered backgrounds.
Integrated Shape-NWD: The first application of Shape-NWD to both the regression loss and Hungarian matching in a DETR-style framework for tiny objects, providing smoother geometric supervision and more stable localization.
Small-Object-Friendly Architecture: The introduction of DLK-GateNet and a P2-augmented PCMF neck to preserve high-resolution cues, forming a robust foundation for prior-guided query initialization.

4. Experimental Results

The method was evaluated on three public benchmarks: DUT-AntiUAV (visible, ground-to-air), DetFly (visible, air-to-air), and CVPR2023 Anti-UAV (infrared).

Performance Gains: Compared to the RT-DETR-R50 baseline, SPQ-DETR achieved significant improvements:
- DUT-AntiUAV: +2.1 AP50, +2.9 AP50–95, +2.5 Recall.
- DetFly: +2.7 AP50, +4.3 AP50–95, +3.1 Recall.
- CVPR2023 Anti-UAV: +2.3 AP50, +2.0 AP50–95, +2.0 Recall.
Efficiency: Despite the architectural enhancements, the model maintains a moderate parameter count (24.7M vs. 41.9M for RT-DETR) and achieves an inference speed of ~30.5 FPS on an RTX 2080 Ti, suitable for practical deployment.
Ablation Studies:
- PCMF and DLK-GateNet improved feature representation, directly benefiting Recall.
- Shape-NWD significantly improved localization stability (AP50–95) and assignment quality.
- SPQ mechanism was the primary driver for reducing missed detections (Recall) by ensuring the decoder attends to high-resolution candidates.
- Using both $P2$ and $P3$ for prior mining and applying a level bias ( $\alpha < 1$ ) yielded optimal results.

5. Significance and Claims

The paper claims that SPQ-DETR addresses the specific failure modes of generic object detectors in anti-UAV scenarios: missed detections due to feature dilution and unstable localization due to geometric metric sensitivity.

Robustness: The method demonstrates consistent improvements across visible and infrared modalities, suggesting that the combination of high-resolution preservation and geometry-aware supervision is modality-agnostic and effective for "few-pixel" targets.
Practicality: By reducing missed detections and improving localization stability without a prohibitive computational cost, SPQ-DETR offers a viable solution for real-world low-altitude security systems where false negatives are costly.
Limitations: The authors modestly acknowledge that the method may still struggle with point-like distractors (e.g., birds, noise) that generate high-confidence responses in high-resolution maps, and performance may degrade under extreme adverse conditions (severe motion blur, over-exposure) where the initial score map becomes unreliable.

In conclusion, SPQ-DETR presents a coordinated design that jointly enhances representation, query initialization, and geometric supervision to tackle the unique challenges of long-range tiny-UAV detection.