TL;DR
Researchers have developed a lightweight, reference-free framework specifically designed to detect and correct "hallucinations"—descriptions of objects or actions not actually present—in aerial image captions. By framing detection as a generative task rather than a simple binary classification, the system achieves significant improvements in factual consistency while remaining fast enough for real-time deployment on drone hardware.
What Problem Does This Paper Address?
Multimodal Large Language Models (MLLMs) are increasingly used to automatically describe aerial imagery captured by drones for applications like disaster response and surveillance. However, these models are notorious for "hallucinating" content that isn't in the image.
This problem is particularly acute in the aerial domain. Unlike standard ground-level photos, aerial images feature extreme viewpoint variations, tiny objects, and significant class imbalances (e.g., thousands of cars but very few rare vehicles). Existing detection methods often rely on massive, computationally expensive models like GPT-4—which actually perform poorly on specialized aerial data—or treat detection as a rigid "yes/no" classification task that fails to capture the nuance and uncertainty of complex visual scenes.
Methods
The authors propose a two-stage, generator-agnostic framework that works with any captioning model.
- Generative Hallucination Detection: Instead of using a traditional classifier, the authors fine-tuned lightweight Multimodal Small Language Models (MSLMs), specifically InternVL3 and Qwen2.5-VL. Given an image and a candidate description, the model is prompted to directly generate a list of hallucinated words or phrases. This "generative" approach allows the model to use its internal reasoning capabilities to pinpoint specific semantic inconsistencies.
- Refinement Module: Once hallucinations are identified, a refinement module (using models like Gemma-4 or Qwen3.5) takes the original image, the flawed description, and the list of detected errors to produce a corrected caption. This module can operate in a "multimodal" mode (using the image for grounding) or a "text-only" mode (editing based on linguistic context).
Datasets
The study primarily utilizes the Labelled Illusion Dataset (LID), which is currently the only publicly available aerial imagery benchmark featuring manual, fine-grained annotations of hallucinated words and spans. To test generalization, the authors also evaluated their models on the HalLoc-Caption dataset, which consists of general-purpose (non-aerial) imagery.
Key Insights and Ideas
- Generative > Discriminative: One of the most important findings is that framing detection as a generation task (asking the model "which words are wrong?") outperforms the standard classification task (asking "is this caption wrong?"). This allows the model to leverage its native instruction-tuning and reasoning strengths.
- Domain Specificity Matters: The paper demonstrates that even powerful proprietary models like GPT-4 struggle with aerial imagery without specific adaptation. Fine-tuning smaller, specialized models on aerial data yields much better results for this niche.
- Confidence as a Proxy for Quality: Because the detector is generative, it naturally produces token-level probabilities. The authors found a strong correlation between the model's "confidence" in its detection and the actual quality of the subsequent correction.
- Efficiency for the Edge: By using MSLMs (1B to 8B parameters), the framework achieves sub-second latency, making it feasible for deployment on embedded AI systems like the NVIDIA Jetson AGX Orin often found on high-end UAVs.
Key Results
- Detection Performance: The best-performing detector (InternVL3-8B) achieved an F1 score of 79.31, significantly outperforming zero-shot baselines and previous classification-based methods like HalLocalizer.
- Hallucination Reduction: The refinement module reduced object-level hallucinations (measured by the CHAIR metric) by up to 60.7%.
- Speed: The InternVL3-1B variant achieved an incredibly fast inference time of 0.064 seconds per sample, with the entire end-to-end pipeline running comfortably under one second per image.
- Human Validation: In a human study, 80% of the refined captions were judged to be less hallucinated than the originals, and 73% were deemed more useful overall.
Limitations and Future Work
While the framework is highly effective, the authors noted that visually complex objects with high intra-class variability (like fountains or specific types of trucks) still pose challenges for fine-grained detection. Additionally, the model's performance dips slightly when a single caption contains multiple different hallucinations.
For future work, the authors plan to extend this framework to video-based UAV captioning, where maintaining temporal consistency across frames introduces a new layer of complexity to the hallucination problem.
Summary updated