RIS-LAD: Low-Altitude Drone Segmentation

• RIS-LAD is a benchmark for low-altitude drone referring image segmentation, defined by 13,871 annotated image-text-mask triplets from realistic urban aerial views.
• It introduces SAARN—a Semantic-Aware Adaptive Reasoning Network that employs CDLE and ARFM to mitigate category drift and object drift in multi-view, small-object scenarios.
• The work highlights unique challenges such as oblique viewpoints, dense same-class instances, clutter, and nighttime scenes, establishing a rigorous testbed for segmentation models.

RIS-LAD is a benchmark-and-model line of work for Referring Low-Altitude Drone Image Segmentation (RLADIS), a setting in which a model must predict a pixel-level mask for the object referred to by a natural-language expression in low-altitude drone imagery. In the formulation introduced by the benchmark paper, RIS-LAD is presented as the first fine-grained RIS benchmark tailored for low-altitude drone scenarios, comprising 13,871 carefully annotated image-text-mask triplets collected from realistic drone footage, together with the Semantic-Aware Adaptive Reasoning Network (SAARN) as a task-specific model. Its central premise is that low-altitude drone views are not a minor variant of natural-image RIS or high-altitude remote-sensing RIS, because they combine diverse viewpoints, small objects, clutter, and dense same-class instances, which in turn induce the failure modes termed category drift and object drift (Ye et al., 28 Jul 2025).

1. Task definition and domain specificity

Referring Image Segmentation (RIS) is the task of predicting a pixel-level mask for the object specified by a referring expression. RIS-LAD specializes this task to low-altitude drone imagery, which the paper denotes as RLADIS. The intended distinction is not merely one of image source. The authors argue that existing datasets and methods are typically designed for high-altitude and static-view imagery, whereas low-altitude drone scenes are characterized by oblique views, stronger viewpoint variation, many small objects, dense same-class instances, cluttered backgrounds, and illumination variation. This directly affects both visual grounding and mask prediction (Ye et al., 28 Jul 2025).

The paper frames two domain-specific failure modes. Category drift occurs when the target object is so small that the model shifts attention toward a larger or more salient object that is semantically similar or category-related. Object drift occurs when multiple instances of the same category are densely packed and the model confuses the intended object with another nearby same-class instance. The distinction is methodologically important: category drift is a failure of early category grounding, whereas object drift is a failure of instance disambiguation after category selection.

A common misconception is that low-altitude drone RIS can be treated as a straightforward extension of either conventional RIS on natural images or Referring Remote Sensing Image Segmentation (RRSIS). The RIS-LAD paper explicitly rejects that equivalence. Its argument is architectural and geometric: natural-image RIS is largely human-centric, RRSIS is largely fixed-view and top-down, while low-altitude drone imagery introduces viewpoint-dependent language, small-object dominance, and crowded same-class scenes that alter both the visual statistics and the semantics of referring expressions.

2. Benchmark construction and dataset properties

RIS-LAD is built from CODrone imagery and is designed to foreground the small-object, multi-viewpoint, urban characteristics of low-altitude drone scenes. The dataset is split in a 7:1:2 ratio for training, validation, and testing, and includes 8 representative object categories: people, car, motor, bicycle, tricycle, truck, bus, and boat. The paper also states that over 90% of annotated instances have a mask coverage ratio below 0.1, making small-object grounding a defining property rather than an incidental difficulty (Ye et al., 28 Jul 2025).

Aspect	RIS-LAD
Core size	13,871 image-text-mask triplets
Split	7:1:2
Image source	CODrone
Patch size	1080 × 1080
Viewpoint	30° to 60°
Categories	8
Nighttime scenes	Present

The construction pipeline is semi-automatic. High-resolution images are split into $1080 \times 1080$ tiles, followed by manual screening to retain tiles that best represent low-altitude drone characteristics. For mask generation, the pipeline uses SAM-2 with the source dataset’s oriented bounding boxes as prompts. For language annotation, the multimodal LLM Qwen2.5-VL receives two inputs: the cropped target object, upsampled for visual clarity, and the full image with a red bounding box marker indicating the target location. The resulting annotations are then carefully verified and refined by humans to maintain semantic consistency among image, referring expression, and mask.

The dataset comparison in the paper emphasizes three properties absent from the listed prior RRSIS datasets: low-altitude drone imagery rather than Google Earth, GF-2, JL-1, or helicopter imagery; oblique $30^\circ$–$60^\circ$ viewpoints rather than fixed views; and the presence of nighttime scenes. The paper does not provide a numerical breakdown of nighttime samples, vocabulary size, average expression length, or the number of original videos and source images. That omission matters for secondary analyses, but it does not affect the central claim that RIS-LAD is structured around a different visual-linguistic regime from prior RIS and RRSIS benchmarks.

3. SAARN architecture and semantic routing

The model proposed with RIS-LAD is SAARN, the Semantic-Aware Adaptive Reasoning Network. SAARN is built on RMSIN and uses a pretrained Swin Transformer as the visual backbone and BERT-base as the language encoder. Its central design principle is that language should not be injected uniformly across the network. Instead, the referring expression is decomposed into three semantic components: global linguistic features $l$, class-level linguistic features $c$, and descriptive linguistic features $d$. This decomposition is then routed to different stages according to the role each component is meant to play (Ye et al., 28 Jul 2025).

The visual features are organized as

$X^f = \{x_1^f, x_2^f, x_3^f, x_4^f\},$

and after pyramid pooling and alignment they become

$$X^{fp} \in \mathbb{R}^{B \times C' \times H' \times W'}.$$

These aligned multi-scale features are processed by two task-specific modules: Category-Dominated Linguistic Enhancement (CDLE) and the Adaptive Reasoning Fusion Module (ARFM).

CDLE is inserted into each stage of the visual encoder and is intended to address category drift. Its logic is sequential. The encoder first aligns visual features with the category token rather than the full expression. Only after that category-level grounding does it inject global sentence information. The paper’s stated purpose is to avoid the situation in which color, relation, or contextual phrases dominate too early and shift the model toward a larger but incorrect object. In the authors’ interpretation, category semantics should stabilize early grounding, while richer linguistic cues should refine it only after category alignment has been established.

ARFM targets object drift. It performs branch-wise cross-modal reasoning using the three semantic branches $l$, $d$, and $30^\circ$0, and then uses a gate to weight them adaptively across scales. The paper defines the gate as

$30^\circ$1

The fused reasoning feature is then

$30^\circ$2

This means that global sentence semantics, descriptive details, and class-level guidance are not trusted equally everywhere. The network learns when each type of cue is most useful.

The paper notes two notation inconsistencies in the architecture description. In CDLE, the global-language attention is written first as $30^\circ$3 and then used as $30^\circ$4. In ARFM, the value projection is written as $30^\circ$5 while the attention operator uses $30^\circ$6. These do not alter the intended semantics of the modules, but they are part of the published technical presentation and are relevant for exact reproduction.

4. Optimization protocol and empirical results

SAARN is trained with AdamW, an initial learning rate of $30^\circ$7, weight decay $30^\circ$8, polynomial decay, 50 epochs, and 4 NVIDIA RTX 3080 GPUs with batch size 2 per GPU. The paper reports the visual backbone as a pretrained Swin Transformer and the language encoder as BERT-base, but it does not provide explicit training-loss formulas, nor does it specify whether BCE, Dice, focal, Lovász, or auxiliary losses inherited from RMSIN are used (Ye et al., 28 Jul 2025).

Evaluation uses Precision@X for $30^\circ$9, together with oIoU and mIoU, which the paper treats as the primary metrics. SAARN is compared against both RIS and RRSIS baselines, including LAVT, ASDA, VATEX, LGCE, FIANet, RMSIN, RSRefSeg, and CADFormer.

Split	oIoU	mIoU
Validation	51.54	44.30
Test	49.60	41.67

On the validation set, SAARN attains [email protected] 47.06, [email protected] 43.12, [email protected] 38.25, [email protected] 31.81, [email protected] 19.27, oIoU 51.54, and mIoU 44.30. On the test set, it attains [email protected] 45.02, [email protected] 39.34, [email protected] 33.37, [email protected] 26.24, [email protected] 15.31, oIoU 49.60, and mIoU 41.67. Relative to RMSIN, the paper reports improvements from 50.17/42.08 to 51.54/44.30 on validation and from 48.82/39.60 to 49.60/41.67 on test for oIoU/mIoU. Relative to RSRefSeg, SAARN also improves the primary metrics, although RSRefSeg is stronger on some stricter precision thresholds such as [email protected].

This produces one of the paper’s more nuanced interpretive points. RSRefSeg attains higher [email protected] on both validation and test, but the authors argue that it often predicts overly large masks, which can inflate strict-threshold precision while not reflecting accurate localization quality. They therefore emphasize oIoU and mIoU as the more reliable indicators. This is less a controversy than a metric-selection warning: different score families privilege different failure modes, and the paper explicitly prefers overlap-based metrics for this task.

The broader empirical message is that RIS-LAD is difficult even for strong methods. For example, VATEX reaches only 24.83/20.32 on validation and 24.27/18.53 on test in oIoU/mIoU, and even the strongest methods remain around 51.54 oIoU / 44.30 mIoU on validation. The benchmark is therefore presented not as near-solved, but as a low-altitude drone grounding problem with substantial remaining headroom.

5. Ablations, failure-mode analysis, and limitations

The ablation studies are organized around the two modules and the three semantic branches. Starting from the baseline configuration without CDLE or ARFM, the paper reports 49.77/42.08 on validation and 48.32/39.60 on test for oIoU/mIoU. With ARFM only, performance becomes 51.31/43.97 on validation and 49.70/40.68 on test. With CDLE only, it becomes 49.82/43.31 on validation and 49.28/41.02 on test. With both modules, it reaches 51.54/44.30 on validation and 49.60/41.67 on test (Ye et al., 28 Jul 2025).

The interpretation offered by the paper is structured. CDLE primarily improves semantic correctness and instance precision by reinforcing category semantics early in the encoder, which is consistent with its proposed role in reducing category drift. ARFM contributes more strongly to oIoU, which the authors interpret as improved localization in dense scenes through adaptive multi-scale reasoning, consistent with its role in reducing object drift.

The semantic-branch ablation is particularly informative. Starting from the CDLE-only configuration, adding global language $60^\circ$0 improves results modestly. Adding $60^\circ$1 without $60^\circ$2 slightly degrades performance relative to $60^\circ$3 alone, which the paper interprets as evidence that descriptive cues can introduce interference when class guidance is absent. Adding $60^\circ$4 gives a strong gain, and using all $60^\circ$5 gives the best overall result. The central claim is therefore not only that semantic decomposition helps, but that descriptive features are useful only when anchored by class-level features.

Qualitative analysis follows the same logic. On tiny-object examples, SAARN is reported to better avoid category drift; in high-density scenarios, it better avoids object drift than RMSIN and RSRefSeg. The paper specifically notes that RSRefSeg often predicts masks that are too large and cover irrelevant neighbors, while other baselines show more obvious confusion among clustered same-class instances.

Several limitations remain explicit or implicit. The paper does not provide exact loss formulas. It does not report cross-dataset transfer, other-dataset evaluation, or generalization studies beyond RIS-LAD. It does not provide separate quantitative breakdowns by object size, daytime versus nighttime, sparse versus dense scenes, or viewpoint bins. It also underreports some reproducibility details, including input resize or crop resolution, data augmentation, warmup schedule, and test-time procedures. These are not peripheral omissions for a benchmark paper, because they delimit how much of the reported gain can be attributed unambiguously to the semantic-routing design rather than to inherited or unstated training choices.

6. Terminology, scope, and neighboring uses of “RIS-LAD”

Within the benchmark paper, RIS-LAD denotes a low-altitude drone benchmark and model for Referring Image Segmentation. In adjacent literature, however, the same string is also used informally for RIS-assisted localization and detection in the reconfigurable-intelligent-surface sensing community. For example, recent works discuss “RIS-LAD” in connection with RIS-aided localization systems, passive radar, or dual-RIS detection support rather than vision-language segmentation (Wu et al., 2024, Bazzi et al., 2022, Liaquat et al., 15 Jul 2025).

This suggests an important terminological caution. RIS-LAD in the low-altitude drone segmentation literature is not an extension of reconfigurable-intelligent-surface sensing, despite the identical acronym. Its “RIS” stands for Referring Image Segmentation, and its “LAD” stands for Low-Altitude Drone. By contrast, in the sensing literature the same label is used as shorthand for localization-and-detection problems involving reconfigurable intelligent surfaces. Conflating the two leads to a category error: one is a multimodal grounding benchmark with a segmentation model, the other is a family of wireless sensing formulations concerned with SNR, geometry, coverage, or target estimation.

In its own domain, RIS-LAD’s significance is more precise. It establishes RLADIS as a distinct task, defines a benchmark tailored to oblique low-altitude drone imagery, and proposes a model whose design is explicitly keyed to the benchmark’s two principal failure modes. Its broader methodological lesson is that, in hard multimodal grounding settings, semantic decomposition and staged fusion can outperform uniform language injection. That principle is the most transferable part of the work, even beyond low-altitude drone referring segmentation.