[Paper Review] Leveraging Explainability for Comprehending Referring Expressions in the Real World

Fethiye Irmak Doğan, Gaspar I. Melsion+1 more|arXiv (Cornell University)|Jul 12, 2021

Multimodal Machine Learning Applications4 citations

TL;DR

This paper proposes a novel method for comprehending referring expressions in real-world robotics by leveraging explainability from vision-language models, specifically using Grad-CAM activations and K-means clustering to identify salient image regions without relying on pre-trained object detectors. The approach outperforms state-of-the-art models in ambiguous and open-vocabulary scenarios, particularly for uncommon or undetected objects.

ABSTRACT

For effective human-robot collaboration, it is crucial for robots to understand requests from users and ask reasonable follow-up questions when there are ambiguities. While comprehending the users' object descriptions in the requests, existing studies have focused on this challenge for limited object categories that can be detected or localized with existing object detection and localization modules. On the other hand, in the wild, it is impossible to limit the object categories that can be encountered during the interaction. To understand described objects and resolve ambiguities in the wild, for the first time, we suggest a method by leveraging explainability. Our method focuses on the active regions of a scene to find the described objects without putting the previous constraints on object categories and natural language instructions. We evaluate our method in varied real-world images and observe that the regions suggested by our method can help resolve ambiguities. When we compare our method with a state-of-the-art baseline, we show that our method performs better in scenes with ambiguous objects which cannot be recognized by existing object detectors.

Motivation & Objective

• To address the limitation of existing referring expression comprehension methods that depend on pre-trained object detectors, restricting them to known object categories.
• To enable robots to identify relevant image regions for ambiguous or uncommon object descriptions in real-world environments where object detectors may fail.
• To improve human-robot collaboration by allowing robots to generate targeted, efficient clarification questions based on salient visual regions.
• To explore the use of model explainability (via Grad-CAM) as a mechanism for grounding language understanding in visual context without supervision on object categories.

Proposed method

• The method uses Grad-CAM to generate class activation maps from a pre-trained image captioning model, identifying visually salient regions for a given referring expression.
• These activation maps are processed to extract high-activation regions that correspond to the described object in the scene.
• K-means clustering is applied to the activation maps to group salient pixels into coherent active clusters representing candidate regions.
• The resulting active clusters are proposed as the robot’s attention targets for follow-up clarification questions.
• The approach does not require object detection or localization modules, making it applicable to any object category, including rare or unseen ones.
• The system is evaluated on real-world images with ambiguous or uncommon objects, comparing performance against MAttNet, a state-of-the-art referring expression model.

Experimental results

Research questions

• RQ1 Can model explainability be used to identify relevant visual regions for referring expressions without relying on object detection?
• RQ2 How does the performance of an explainability-based method compare to state-of-the-art referring expression models in scenes with ambiguous or uncommon objects?
• RQ3 Can the proposed method generate effective clarification questions by focusing on salient image regions rather than full scenes?
• RQ4 To what extent can the method handle object categories not covered by existing object detectors?

Key findings

• The proposed method outperforms MAttNet in scenes with multiple similar distractors, such as multiple fish, where the baseline fails to distinguish between candidates.
• The method successfully identifies relevant regions for uncommon or undetected objects—such as an artichoke or papaya—where object detectors fail, demonstrating robustness in open-vocabulary settings.
• Qualitative analysis shows that the method focuses on semantically relevant image regions (e.g., pants in a description mentioning 'the man in pants'), indicating effective grounding of language in vision.
• Active clusters generated by the method are effective for formulating targeted clarification questions, such as 'Do you mean the fish on the left or the right?', reducing the need for full request repetition.
• The method enables robots to handle ambiguous descriptions by focusing on salient, contextually relevant image regions, improving interaction efficiency.