PaperlayerView exact revision on arXiv

Source-linked AI summary

Aerial-R1: Reinforcing Aerial Reasoning and Segmentation in MLLMs

Jisheng Dang, Huaiyuan Peng, Yixuan Wang, Kun Wang, Zhang Li, Tat-Seng Chua

arXiv:2610.33056v1cs.CVcs.AI

TL;DR

Aerial MLLMs can hallucinate under visual ambiguity, while SFT teaches target identities but not when to predict. Aerial-R1 uses H-GAWR to reason across resolutions; VSAI-Ref experiments report state-of-the-art precision on tiny objects and an incentivized refusal boundary when evidence is insufficient.

  • Problem

    SFT teaches target identities but not when to predict, while trustworthy visual perception should remain stable across resolutions.

    Source: §I · p. 1, A. Visual Consistency Motivation · p. 14

  • Method

    Aerial-R1 uses Hybrid-View Group Advantage-Weighted Regression to optimize non-autoregressive mask decoders and reason about visual consistency across resolutions.

    Source: §VI · p. 12

  • Results

    Experiments on VSAI-Ref report state-of-the-art precision on tiny objects and an explicitly incentivized safety boundary for refusing predictions when visual evidence is insufficient.

    Source: §VI · p. 12

  • Takeaways & Limitations

    The authors conclude that advantage-weighted visual reasoning may support trustworthy aerial agents that refuse predictions when evidence is insufficient.

    Source: §VI · p. 12

    Under severe occlusion, vehicles more than 80% obscured can trigger empty-mask refusals, while fine-grained vehicle and color distinctions remain challenging in low-resolution views.

    Source: G. Limitations and Failure Case Analysis · p. 12

Abstract

from arXiv · show

While Multimodal Large Language Models (MLLMs) have revolutionized general visual perception, existing methods suffer from severe hallucination in the aerial domain, where extreme scale variations and visual ambiguity abound. Existing Supervised Fine-Tuning (SFT) paradigms force models to mimic ground truth without modeling uncertainty, often resulting in blind guessing on tiny or unrecognizable objects in aerial scenes. In this paper, we present Aerial-R1, the first aerial MLLM framework to reinforce visual reasoning capabilities for aerial reasoning and segmentation. Drawing inspiration from the reasoning-incentivizing mechanisms of DeepSeek-R1, we find that trustworthy perception stems from consistency across views, establishing a visual parallel to the Chain-of-Thought in logic. To effectively adapt policy optimization for non-autoregressive mask generation, we introduce Hybrid-View Group Advantage-Weighted Regression (H-GAWR). This strategy constructs dynamic groups of high-resolution and low-resolution views, seamlessly bridging group-relative advantage estimation with advantage-weighted policy extraction. By enforcing consensus on clear inputs and strictly penalizing hallucinations on ambiguous ones, Aerial-R1 develops an explicitly incentivized refusal mechanism. Extensive experiments on VSAI-Ref, the first aerial MLLM benchmark dedicated to jointly evaluating dense reasoning and hallucination robustness, demonstrate that Aerial-R1 significantly outperforms SFT baselines in segmentation precision while substantially mitigating fatal hallucinations in low-information scenarios, paving the way for more reliable and trustworthy aerial agents. Codes are available on https://github.com/phy709/Aerial-R1.

I. INTRODUCTION

Aerial imagery’s tiny, ambiguous targets expose limits in SFT-based segmentation, which can produce unstable boundaries and hallucinated masks. Aerial-R1 uses consistency across image resolutions and an incentivized refusal mechanism to pursue more trustworthy aerial perception.

A. Referring Image Segmentation in Aerial Imagery

  • SFT-based aerial models can produce coarse, inconsistent boundaries and hallucinated masks for tiny or visually ambiguous targets.The paper identifies extreme scale variation, dense tiny objects, and clutter as challenges, and notes that prior mitigation can require costly preference annotations or external detectors.Source: §I · p. 1, A. Referring Image Segmentation in… · p. 3

I. INTRODUCTION

  • The paper argues that SFT teaches target mappings but not when to predict, motivating visual consistency across resolutions as a signal of trustworthy perception.Aerial-R1 treats inconsistency as uncertainty and calls for silence when visual evidence is corrupted.Source: §I · pp. 1–2
  • Aerial-R1 introduces H-GAWR, which combines hybrid-view group sampling with Advantage-Weighted Regression to optimize precision and safety without manual negative annotations.It enforces consensus on high-resolution views and penalizes hallucinations on low-resolution views to incentivize refusal.Source: §I · p. 2
  • VSAI-Ref is introduced as an aerial benchmark designed to evaluate dense visual grounding and cognitive safety, including a dedicated refusal testbed.The paper describes it as the first comprehensive multimodal aerial benchmark to integrate that testbed.Source: §I · p. 2

B. Hallucination Mitigation in MLLMs

The paper frames aerial hallucination mitigation as a mismatch between general mitigation approaches and dense aerial segmentation, and motivates stable advantage-weighted learning for non-autoregressive masks. Aerial-R1 combines that learning strategy with a multimodal segmentation architecture.

B. Hallucination Mitigation in MLLMs

  • External-detector correction and preference-based RLHF face structural difficulties in dense aerial scenes, where tiny, ambiguous targets challenge general-domain detectors.The paper therefore benchmarks against Sa2VA as a representative MLLM-SAM hybrid SFT model.Source: B. Hallucination Mitigation in MLLMs · p. 3

C. Reinforcement Learning for Reasoning

  • AWR reframes policy updates as advantage-weighted supervised learning, while group-relative consensus sampling draws on GRPO to support stable learning for non-autoregressive mask decoders.The paper notes that direct token-level policy gradients are problematic for high-dimensional continuous masks.Source: C. Reinforcement Learning for Reasoning · p. 3

III. METHODOLOGY

  • Aerial-R1’s H-GAWR is presented as a self-supervised mechanism for boundary consistency and suppressing hallucinations in ambiguous scenarios.The reward architecture separates segmentation accuracy and hybrid consistency from penalizing hallucinations or rewarding silence on ambiguous inputs.Source: §III · p. 3, Figure 2 · p. 4

A. Preliminaries and Architecture

  • Aerial-R1 uses Sa2VA with an InternVL-Chat-V1.5 backbone and LoRA adaptation, while SAM-2 generates the final mask from a grounding prompt.The [SEG] token supplies a hidden state that is transformed into a dynamic prompt for SAM-2’s non-autoregressive decoder.Source: A. Preliminaries and Architecture · pp. 3–4

B. Hybrid-View Group Construction

Hybrid-View Group Sampling addresses redundant exploration in ambiguous aerial imagery by pairing pristine views with degraded variants. Gaussian blur is used to model information loss in a way the paper relates to aerial imaging conditions.

B. Hybrid-View Group Construction

  • Hybrid-View Group Sampling is proposed to counter premature mode collapse and redundant exploration in ambiguous aerial mask spaces by decoupling semantic priors from structural evidence.Each training image and referring text yields a diverse group of views.Source: B. Hybrid-View Group Construction · p. 4
  • Each group combines pristine high-resolution copies retaining the target mask with degraded low-resolution variants assigned an empty ground truth.The low-resolution views represent an information-depleted state.Source: B. Hybrid-View Group Construction · p. 4
  • The degradation operator is isotropic Gaussian blur, implemented as convolution with a 15×15 kernel and σ = 5.0.The paper motivates blur over pixelation or random noise as a better fit to atmospheric scattering, sensor defocusing, and optical diffraction.Source: B. Hybrid-View Group Construction · p. 4
  • Blur attenuates high-frequency boundaries while preserving low-frequency global cues, creating cases where a target may be suggested but lacks sufficient spatial support.The paper uses this contrast to represent uncertainty and penalizes hallucinated masks in information-depleted scenarios.Source: B. Hybrid-View Group Construction · p. 5

C. Hybrid-View Group Advantage-Weighted Regression (H-GAWR)

H-GAWR adapts group-relative optimization to dense mask prediction by combining high- and low-resolution views with separate precision and refusal rewards. Clear views reward target accuracy and group consensus, while unrecognizable views penalize hallucinated foreground masks.

C. Hybrid-View Group Advantage-Weighted Regression (H-GAWR)

  • H-GAWR combines group-relative advantage estimation with Advantage-Weighted Regression, avoiding external critics for dense, non-autoregressive mask prediction.The approach addresses instability and intractability of applying standard policy optimization directly to continuous mask decoders.Source: C. Hybrid-View Group… · p. 5

1) Dual-Branch Reward Mechanism:

  • The reward uses separate objectives: high-resolution views emphasize boundary precision, while low-resolution views calibrate uncertainty and refusal.Low-resolution refusal rewards apply when the target is fully unrecognizable and the ground-truth mask is empty.Source: 1) Dual-Branch Reward Mechanism · p. 5
  • For valid high-resolution targets, candidate masks are rewarded for both overlap with ground truth and agreement with a group consensus mask.The composite reward is IoU(Mi, Mgt) + β · IoU(Mi, Mcons), with β scaling the consensus reward.Source: 1) Dual-Branch Reward Mechanism · p. 5
  • The consensus mask is derived by thresholding a pixel-wise mean map of the group’s continuous logits.This consensus target is used to encourage consistency across views.Source: 1) Dual-Branch Reward Mechanism · p. 5
  • For unrecognizable low-resolution views, the refusal reward favors masks below a 50-pixel foreground threshold and penalizes larger predictions.The threshold tolerates negligible isolated activations; the branch shifts from localization to hallucination suppression.Source: 1) Dual-Branch Reward Mechanism · p. 5

2) Contrastive Advantage Estimation:

H-GAWR compares candidate rewards within each group to estimate relative advantages, then uses positive-advantage masks for weighted structural supervision. If no candidate has positive advantage, it falls back to unweighted mask loss to preserve gradient flow.

2) Contrastive Advantage Estimation:

3) Advantage-Weighted Policy Extraction:

  • Standard policy gradients are unsuitable for the dense, non-autoregressive SAM-2 mask decoder because high-dimensional pixel outputs have extreme variance.The method therefore propagates reinforcement signals through advantage-weighted structural supervision rather than explicit policy gradients.Source: 3) Advantage-Weighted Policy Extraction · p. 5
  • Only better-than-average trajectories with positive advantages are retained, and their advantages weight the structural mask objective alongside LLM log-likelihood.The mask loss balances pixel-level classification and region overlap using 2.0 · LCE + 0.5 · LDice.Source: 3) Advantage-Weighted Policy Extraction · p. 5
  • When no positive-advantage trajectories exist, training falls back to the unweighted mean mask loss to maintain gradient flow.This fallback is described for the rare case of severe mode collapse.Source: 3) Advantage-Weighted Policy Extraction · p. 5

4) Theoretical Justification of H-GAWR:

The theoretical justification frames H-GAWR as projecting an optimal policy onto the neural policy through weighted maximum-likelihood estimation. A separate exploration mechanism perturbs grounding embeddings with scaled noise during training.

4) Theoretical Justification of H-GAWR:

  • The optimal policy is projected onto the neural policy by minimizing their KL divergence, equivalent to weighted maximum-likelihood estimation.The derivation frames the search using a KL constraint and a positive Lagrange multiplier.Source: 4) Theoretical Justification of H-GAWR · p. 6

5) Continuous Embedding Exploration:

  • Continuous Temperature Sampling perturbs grounding embeddings with scaled stochastic noise to explore the mask decoder’s continuous geometric space.The noise is standard Gaussian, and γ controls exploration intensity; the perturbation is intended to improve robustness to minor semantic fluctuations and prevent suboptimal geometric-mode collapse.Source: 5) Continuous Embedding Exploration · p. 6

IV. THE VSAI-REF BENCHMARK

VSAI-Ref is designed to assess both dense aerial grounding and hallucination robustness, addressing the positive-only, object-centric scope of established aerial datasets. Its annotation pipeline combines referring segmentation, visual question answering, scene descriptions, and adversarial hard negatives.

IV. THE VSAI-REF BENCHMARK

  • VSAI-Ref jointly evaluates dense grounding precision and hallucination robustness, unlike positive-only, object-centric aerial benchmarks.It retains the original VSAI dataset’s high-altitude complexity, sensor characteristics, scale variation, and environmental diversity.Source: §IV · p. 6

A. Pipeline Construction via MLLMs

  • The annotation pipeline uses Qwen-VL-Max to generate referring expressions, visual question-answer pairs, and scene descriptions from VSAI imagery.The benchmark includes 214 scene descriptions covering global context, spatial distributions, and inter-object relationships.Source: A. Pipeline Construction via MLLMs · p. 7

B. The Hard Negative Subset

  • The hard-negative test set contains 832 vehicle-free images paired with context-aware deceptive vehicle descriptions to test explicit refusal.Queries are selected from a manually curated pool of 13 error-prone descriptions.Source: B. The Hard Negative Subset · p. 7

C. Dataset Statistics

  • The benchmark includes 5,372 positive referring-segmentation samples and 832 hard negatives, with negatives reserved for testing.Figure 3 depicts the proportions of referring segmentation, VQA, scene-description, and hard-negative samples; Table I summarizes split-level counts.Source: C. Dataset Statistics · p. 7, Figure 3 · p. 7, Table I · p. 7

V. EXPERIMENTS

Experiments evaluate Aerial-R1 on VSAI-Ref across segmentation accuracy, hallucination refusal, robustness, and component ablations. Results report improved segmentation metrics and refusal rates, with ablations examining the roles of consistency and low-resolution views.

A. Experimental Setup

  • Training uses SFT followed by H-GAWR; hybrid groups increase training compute, but the added complexity is confined to training and adds no inference latency or memory overhead.The group size is G = 8, and H-GAWR converges in a single epoch.Source: A. Experimental Setup · p. 8

B. Main Results

  • Compared with standard SFT, Task 1 Aerial-R1 (Noise) gains 3.16% mIoU and 3.13% F1-Score; Task 2 Aerial-R1 (Blur) reaches 57.53% mIoU and 68.86% Recall (+2.75% over SFT).The low-resolution augmentation baseline lowers Task 1 mIoU from 42.96% to 40.83%.Source: B. Main Results · p. 8

C. Robustness Analysis

  • H-GAWR refuses 61.18% of 832 hard negatives, versus 38.82% for SFT and 52.64% for low-resolution augmentation; visualizations cover spatial scenarios and varying lighting.On hard negatives, SFT produces hallucinated masks while Aerial-R1 outputs empty masks when visual evidence is lacking.Source: C. Robustness Analysis · pp. 8–9, Figure 4 · p. 9, Figure 5 · p. 10

D. Ablation Study

  • Removing Consistency Reward lowers Task 1 mIoU by 1.65% and Recall by 1.54%; removing the Low-Res stream lowers mIoU by 1.48% in Task 1 and 1.04% in Task 2.The full model reaches 57.53% mIoU in Task 2 by integrating consistency and explicitly incentivized uncertainty.Source: D. Ablation Study · pp. 9–10

E. Extension: Generalized Reasoning Evaluation

Aerial-R1 is evaluated for generalized reasoning and training stability, with reported gains on VQA and scene description alongside convergent mask training and a learned refusal boundary.

E. Extension: Generalized Reasoning Evaluation

  • The VQA and holistic scene-description evaluations test whether dense-task alignment preserves general reasoning, and Table V examines component removals across both tasks.The VQA benchmark includes attribute identification and numerical reasoning; Figure 6 illustrates a multi-hop spatial query and a global scene description.Source: E. Extension: Generalized Reasoning… · pp. 10–11, Table V · p. 11, Figure 6 · p. 11
  • Table III reports the best overall mIoU/F1 trade-off across both tasks, while Table IV reports the highest successful refusal rate on 832 hard negatives.Table III compares standard SFT and naive augmentation with blur, pixelation, and noise variants; Table IV compares refusal performance with SFT and naive augmentation.Source: Table III · p. 11, Table IV · p. 11
  • 71.73% VQA accuracy (+1.90% over standard SFT) and CIDEr 53.99 versus 37.88 mark stronger generalized reasoning; naive Sa2VA augmentation degrades ROUGE-L and CIDEr.Table VI reports overall VQA accuracy and scene-description ROUGE-L and CIDEr scores.Source: Table VI · p. 11, E. Extension: Generalized Reasoning… · p. 11

F. Training Dynamics & Stability Analysis

G. Limitations and Failure Case Analysis

Aerial-R1’s extreme-case failures expose a conservative refusal trade-off and remaining fine-grained attribute errors; the conclusion frames cross-resolution reasoning and refusal as central outcomes.

G. Limitations and Failure Case Analysis

  • Severe occlusion can trigger an empty-mask false negative, while similar vehicle categories and dark colors can still produce attribute mismatches from language priors.The refusal tendency is reported for vehicles more than 80% obscured; the attribute challenges occur in low-resolution aerial views.Source: G. Limitations and Failure Case Analysis · p. 12

VI. CONCLUSION

  • The conclusion presents advantage-weighted visual reasoning as a shift from supervised fitting toward trustworthy aerial perception.It describes the approach as a foundation for more trustworthy autonomous aerial agents.Source: §VI · p. 12
  • Aerial-R1 reasons about visual consistency across resolutions, and VSAI-Ref experiments report state-of-the-art tiny-object precision and an explicitly incentivized refusal boundary.The conclusion presents refusal when visual evidence is insufficient as a safety capability.Source: §VI · p. 12

I. ADDITIONAL BACKGROUND AND MOTIVATION

The background motivates cross-resolution consistency and refusal as responses to aerial ambiguity and hallucination, then outlines why existing mitigation and policy-optimization approaches do not directly fit dense aerial segmentation.

A. Visual Consistency Motivation

  • Aerial imagery combines extreme scale variation, tiny densely packed objects, and clutter; the paper argues SFT mimics targets without teaching when to predict, encouraging unsupported answers.The proposed visual analogy is that trustworthy perception should remain stable across resolutions and stay silent when evidence is degraded.Source: A. Visual Consistency Motivation · p. 14

B. Hallucination Mitigation in MLLMs

  • Post-hoc detectors and preference-based alignment face aerial-domain obstacles, motivating degraded views as intrinsic uncertainty proxies for learning refusal without manual negatives or external estimators.The cited constraints include tiny ambiguous targets for external detectors and the difficulty of constructing pixel-level preference trajectories.Source: B. Hallucination Mitigation in MLLMs · p. 14

C. Reinforcement Learning for Reasoning

  • Token-level GRPO is problematic for dense non-autoregressive masks; AWR offers stable advantage-weighted policy extraction, combined with group-relative consensus sampling.The background describes high variance and unstable convergence in high-dimensional pixel spaces as challenges for direct policy gradients.Source: C. Reinforcement Learning for Reasoning · p. 15

A. Preliminaries and Architecture

  • The architecture uses Sa2VA to connect LLM semantic reasoning with SAM-2 localization, with the mask decoder adapted while the vision encoder remains frozen.InternVL-Chat-V1.5 and LoRA provide the cognitive engine and aerial-domain adaptation; the [SEG] token prompts SAM-2 through a projection neck.Source: A. Preliminaries and Architecture · p. 15

B. Physical Motivation of Gaussian Degradation

The method uses Gaussian blur to model aerial visual degradation, then combines hybrid-view rewards and advantage-weighted mask updates with language-model training to promote stable prediction and refusal.

B. Physical Motivation of Gaussian Degradation

  • A 15 × 15 Gaussian blur with σ = 5.0 models atmospheric scattering, defocusing, and diffraction while attenuating high-frequency detail more naturally than pixelation or random noise.The method preserves low-frequency global semantics while weakening object boundaries and fine textures, creating an information-depleted view.Source: B. Physical Motivation of Gaussian… · p. 15

C. Reward and Advantage Details

  • Group consensus aggregates logits from pristine and degraded views, while a 2.0 hallucination penalty versus 1.0 silence reward prioritizes refusal when evidence is compromised.Group-wise normalization compares predictions without a separate critic; hallucination on a low-resolution view can raise the relative advantage of successful high-resolution localization.Source: C. Reward and Advantage Details · pp. 15–16

D. Joint Objective and AWR Interpretation

  • The joint objective preserves language-model loss as a semantic anchor and uses a truncated AWR mask objective that selects positive-advantage trajectories to limit gradient explosion.AWR constrains policy divergence from a reference; in this non-autoregressive setting, policy log-likelihood maximization corresponds to minimizing structural mask loss.Source: D. Joint Objective and AWR… · p. 16

Related papers

Loading 2610.33056v1…