TLDR: GroundSight is a new method that improves Visual Question Answering (VQA) by augmenting Vision-Language Models (VLMs) with text-grounded object localization and de-hallucination. It allows models to generate bounding boxes around relevant objects for focused information retrieval, reducing background noise and improving accuracy. Additionally, it fine-tunes models to say “I don’t know” for uncertain questions, significantly cutting down hallucination rates and boosting the truthfulness of responses. The combined approach leads to more reliable and context-aware VQA.
Visual Question Answering (VQA) systems, which allow machines to answer questions about images, are becoming increasingly sophisticated. However, these systems often struggle with two main issues: accurately identifying the specific object a question refers to within a complex image, and sometimes generating confident but incorrect answers, a phenomenon known as hallucination. A new research paper introduces GroundSight, a novel approach designed to tackle these challenges by enhancing Vision-Language Models (VLMs) with better object grounding and a mechanism to reduce hallucinations.
Current VLMs, while powerful, are limited by the knowledge embedded in their training data. When combined with Retrieval-Augmented Generation (RAG), which fetches external information, the challenge intensifies. It becomes difficult to retrieve knowledge that is relevant to both the text query and the visual content. For instance, if you ask about the cost of an object held in someone’s hand, but the background is cluttered with other items, the VLM needs to know precisely which object to focus on for retrieval. Without this focus, the system might retrieve irrelevant information, leading to incorrect or misleading answers.
Introducing GroundSight: A Smarter Approach to VQA
GroundSight aims to improve VQA in real-world scenarios where ambiguous questions and busy backgrounds can confuse models. The core idea is to enable the model to generate a bounding box around the object most relevant to the question. This targeted approach allows for precise image cropping and focused information retrieval, significantly reducing background noise and improving the alignment between visual and textual cues. This method also plays a crucial role in mitigating hallucinations.
The GroundSight system operates within a Retrieval-Augmented Generation (RAG) framework, incorporating three main components:
- Vision-Language Model (VLM): The system experimented with models like BLIP, QWen, and LLaMA 3.2, ultimately selecting LLaMA 3.2 for its performance.
- Region of Interest Proposer: This module is responsible for pinpointing the object or area in the image that is most relevant to the question.
- Image-Based Information Retriever: This component performs searches using the cropped image regions, either from web sources or local databases.
Beyond these, GroundSight also fine-tunes the VLM to reduce hallucinations in uncertain situations.
Localizing the Region of Interest
A key innovation of GroundSight is its ability to localize the region of interest. Instead of processing an entire image, the model is trained to identify and draw a bounding box around the specific object mentioned in the question. This cropped region is then used for a more focused visual search. The retrieved information is then fed back to the VLM to formulate the final answer.
The researchers explored two main ways to make the RAG agent region-aware:
- Using a Pretrained Localizer: They incorporated a pretrained localizer, such as Grounding DINO, to identify regions of interest. This approach is efficient and effective for visual search.
- Training Vision-Language Models for Grounding: They also investigated training VLMs like BLIP-2 to predict bounding boxes directly. While promising, the final GroundSight solution utilized Grounding DINO due to computational constraints.
Combating Hallucination
Hallucinations, where models generate plausible but factually incorrect answers, are a major concern for VLM reliability. GroundSight addresses this by fine-tuning the model to respond with “I don’t know” when it is uncertain, particularly for question types that typically require external knowledge, such as “who” questions. This strategic abstention significantly reduces the hallucination rate and improves the overall truthfulness of the responses.
An additional de-hallucination strategy involved applying an image search threshold. By setting a CLIP similarity threshold of 0.75, the system filters out irrelevant search results, ensuring that only semantically grounded matches are considered, further preventing misleading information from influencing the model’s answers.
Performance and Key Insights
The GroundSight system was evaluated on a challenging VQA dataset derived from Meta RayBan smart glasses, featuring diverse domains and varying image quality. The evaluation used GPT-4o-mini as a judge to assign a “truthfulness score” based on accuracy, missing rate, and hallucination rate.
The results demonstrated that while localization techniques like using Grounding DINO alone initially led to a slight drop in accuracy, combining it with a “Chain-of-Spot” prompting strategy improved accuracy. However, Chain-of-Spot also increased the hallucination rate, as the model became overly confident. The most significant improvement came from the de-hallucination method, which drastically reduced the hallucination rate from 65.79% to 13.88% and improved the truthfulness score.
The full GroundSight agent, integrating de-hallucination with Grounding DINO and Chain-of-Spot prompting (and an image search threshold), achieved the highest truthfulness score. This highlights a crucial insight: in real-world applications, admitting uncertainty by saying “I don’t know” is often more beneficial than providing a confidently incorrect answer. This balance of grounding, confidence in its own knowledge, and cautiousness when uncertain, leads to more reliable and trustworthy VQA systems.
For more in-depth details, you can read the full research paper here.
Also Read:
- Improving Robot Navigation with Contextual Textual Descriptions in LLMs
- NePTune: A Neuro-Symbolic Framework for Advanced Vision-Language Reasoning
Conclusion
GroundSight offers a practical and effective method for improving VQA performance by integrating text-grounded object localization and a robust de-hallucination strategy. By enabling models to focus on relevant image regions and to express uncertainty when appropriate, GroundSight paves the way for more accurate, context-aware, and trustworthy answers from vision-language models.
