RIS-LAD: Referring Low-Altitude Drone Image Segmentation

RIS-LAD is the first fine-grained Referring Image Segmentation benchmark for low-altitude drone (LAD) scenes, featuring 13,871 image–text–mask triplets.
We introduce SAARN (Semantic-Aware Adaptive Reasoning Network) with CDLE and ARFM to tackle LAD-specific category drift and object drift.

---

🔗 Quick Links

• Paper (arXiv): https://arxiv.org/abs/2507.20920
• Dataset (Google Drive): https://drive.google.com/file/d/1PmtaQH_F0AUoGWgpmDSpPu27E2XSdGd4/view?usp=sharing
• Model & Code: coming soon

---

🧾 Introduction

Low-altitude drones (typically operating below ~200 m) are increasingly deployed in real-world perception systems thanks to their flexibility and cost-effectiveness. However, most existing referring image segmentation (RIS) research focuses on conventional ground-view scenes or high-altitude remote sensing imagery. These settings differ substantially from low-altitude drone (LAD) views, where perspectives are oblique, objects are tiny and densely packed, and illumination varies widely (including night scenes).

To bridge this gap, we present RIS-LAD, a fine-grained benchmark specifically designed for Referring Low-Altitude Drone Image Segmentation (RLADIS). RIS-LAD offers 13,871 carefully verified image–text–mask triplets collected from real LAD footage. Beyond providing data, RIS-LAD formalizes two key failure modes frequently observed under LAD settings:

• Category drift: when tiny targets cause models to latch onto larger, semantically similar objects.
• Object drift: when crowds of same-class instances lead to confusion about which instance the expression refers to.

We further propose SAARN (Semantic-Aware Adaptive Reasoning Network) to tackle these challenges. SAARN introduces:

• CDLE (Category-Dominated Linguistic Enhancement): injects class-level linguistic cues early in the encoder to anchor visual features to the correct category and suppress category drift.
• ARFM (Adaptive Reasoning Fusion Module): performs scale-aware fusion of global (l), class (c), and descriptive (d) text cues, enabling coarse→fine reasoning and disambiguation among dense same-class instances.

RIS-LAD is built with a semi-automatic pipeline that combines high-quality instance masks (prompted SAM-2) and multimodal LLM–generated initial expressions (given cropped instances and location cues), followed by human refinement. Experiments on RIS-LAD show that SAARN achieves state-of-the-art results on core segmentation metrics and yields pronounced gains at stricter localization thresholds (e.g., P@0.9), demonstrating stronger instance-level discrimination under LAD conditions.

---

✨ Highlights

_{Figure 1. RLADIS challenges vs. RRSIS (category/object drift, tiny & dense objects, illumination).}

• Low-altitude & oblique views (≈30–100 m, 30°–60°) → strong perspective change & foreshortening.

• Tiny & dense targets → easy confusion among many same-class instances.

• Variable illumination (incl. night) → distribution shift from standard RS/RIS datasets.

• Category drift: tiny targets bias the model to large, semantically similar objects.

• Object drift: crowded same-class instances hinder precise instance selection.

---

📊 Dataset Characteristics

Dataset	Image Source	Shooting Angle	Nighttime Scene
RefDIOR	Google Earth	fixed	✗
NWPU-Refer	Google Earth	fixed	✗
RISBench	Google Earth, GF-2, JL-1	fixed	✗
RefSegRS	Helicopter (above 1000 m)	fixed	✗
RRSIS-D	Google Earth	fixed	✗
RIS-LAD (ours)	Drone (30–100 m)	30°–60°	✓

---

📈 Benchmark Results

Method	Publication	oIoU (Val)	oIoU (Test)	mIoU (Val)	mIoU (Test)
LAVT	CVPR 2022	44.03	41.97	32.25	30.14
ASDA	MM 2024	38.70	37.53	35.46	33.33
VATEX	WACV 2025	24.83	24.27	20.32	18.53
LGCE	IEEE TGRS 2024	41.72	40.75	27.68	26.17
FIANet	IEEE TGRS 2024	45.24	43.39	39.61	37.44
RMSIN	CVPR 2024	50.17	48.82	42.08	39.60
RSRefSeg	IGARSS 2025	50.04	47.71	43.42	41.16
CADFormer	JSTARS 2025	47.37	46.47	41.36	39.32
SAARN (ours)	—	51.54	49.60	44.30	41.67

---

🖼️ Qualitative Comparisons

_{Qualitative Results. SAARN vs. prior SOTA on tiny-object and dense same-class cases.}

---

🧭 Roadmap

• Release training & evaluation code for SAARN
• Add dataset loaders for popular toolkits
• Provide pretrained weights & full configs
• Add extended qualitative gallery and failure cases

---

📝 Citation

@misc{ye2025risladbenchmarkmodelreferring, 
  title        = {RIS-LAD: A Benchmark and Model for Referring Low-Altitude Drone Image Segmentation}, 
  author       = {Kai Ye and YingShi Luan and Zhudi Chen and Guangyue Meng and Pingyang Dai and Liujuan Cao},
  year         = {2025},
  eprint       = {2507.20920},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2507.20920}
}

---

📫 Contact & License

• Issues or questions: yekai@stu.xmu.edu.cn
• Dataset: research use only (see dataset README)
• Code: license will be provided upon release