TL;DR
This paper introduces the VQA-HAT (Human ATtention) dataset, which maps exactly where humans look when answering visual questions using a novel "deblurring" interface. By comparing these human maps to state-of-the-art deep learning VQA models, the authors demonstrate that current attention networks do not look at the same regions as humans, though more accurate models show a slight improvement in alignment.
What Problem Does This Paper Address?
Visual Question Answering (VQA) requires AI models to understand both text and images to answer open-ended questions (e.g., "What color is the fire hydrant?"). To solve this, many state-of-the-art VQA models employ "attention mechanisms" designed to highlight relevant image regions.
However, these attention maps are typically learned in an unsupervised manner—they are simply intermediate spatial distributions optimized to reduce final answer loss. It has long been unclear whether these networks are actually "reasoning" like humans or just finding fortuitous statistical shortcuts. This paper addresses a fundamental question: Do deep VQA models actually look at the same image regions as humans when answering a question?
Methods
To capture true task-driven human attention without expensive eye-tracking hardware, the authors developed a game-inspired crowdsourcing interface on Amazon Mechanical Turk.
Users are presented with a completely blurred image and a question. They must use a "click-and-drag" coloring motion with their mouse to progressively sharpen the regions of the image necessary to answer the question accurately.
The authors experimented with three interface variants and found that showing the question and the correct answer, then asking users to sharpen as few regions as possible, yielded the highest quality data.
To compare human maps with machine-generated attention maps, the authors utilized Rank Correlation on 14x14 scaled grids. This order-based metric ranks pixels by their attention intensity, making the evaluation invariant to absolute probability values (which can easily be altered by tweaking a model's softmax temperature).
Datasets
- VQA Dataset: The standard free-form, open-ended benchmark used as the foundation for image-question pairs.
- VQA-HAT (Human ATtention) Dataset: The newly introduced dataset containing human attention maps for 58,475 training and 1,374 validation question-image pairs.
Key Insights and Ideas
- Attention is highly task-driven: Human attention changes drastically on the exact same image depending on the question asked. For instance, looking at the floor for "What type is the surface?" versus looking at the racket/player for "Which game is being played?".
- Networks don't see like us: Current unsupervised attention models do not strongly align with human attention, suggesting that their internal "reasoning" is fundamentally different from ours.
- The Center Bias Illusion: Initially, standard task-independent saliency models (which predict where humans look during natural, unguided viewing) correlated better with humans than VQA models did. However, the authors discovered this was an illusion caused by "center bias"—the tendency for both photographers and humans to focus on the center of an image. Once center-biased images were filtered out, the task-independent model's correlation plummeted to zero, while VQA-specific models maintained their performance.
- Performance correlates with alignment: As VQA models become more accurate (e.g., moving from Stacked Attention Networks to Hierarchical Co-Attention Networks), their attention maps become slightly better correlated with human attention.
Key Results
- The baseline for human-to-human agreement (inter-human rank correlation) is 0.623.
- On the full validation set, a task-independent saliency model (Judd et al.) achieved a rank correlation of 0.497, outperforming the best VQA model, HieCoAtt-Q, which scored 0.264 (SAN-2 scored 0.249).
- After removing center-biased maps to isolate truly task-driven attention, the ranking flipped dramatically: HieCoAtt-Q scored 0.114, SAN-2 scored 0.038, and the task-independent Judd saliency dropped to -0.063.
Limitations and Future Work
The authors note that defining a "necessary and sufficient" attention map at the pixel level is fundamentally an ill-posed problem; random pixels can be blacked out and humans can often still guess the answer. They suggest that future work should explore the right semantic space (such as objects or attributes rather than raw pixels) to evaluate attention meaningfully.
Crucially, the introduction of the VQA-HAT dataset provides a benchmark that future researchers can use to transition from unsupervised attention to explicitly supervised attention training.
Summary updated