Human Attention in Visual Question Answering: Do Humans and Deep Networks Look at the Same Regions?

Abhishek Das, Harsh Agrawal, C. Lawrence Zitnick, Devi Parikh, Dhruv Batra

TL;DR

This paper investigates whether current deep learning models for Visual Question Answering (VQA) "look" at the same image regions as humans when answering questions. They introduce a new dataset, VQA-HAT, collected through novel interactive interfaces, and find that state-of-the-art VQA attention models do not align well with human attention patterns, suggesting a disconnect in how these models process visual information for question answering.

What Problem Does This Paper Address?

Visual Question Answering (VQA) systems aim to answer natural language questions about images. A key component in many advanced VQA models is an "attention mechanism," which is designed to mimic human visual attention by allowing the model to focus on specific regions of an image that are most relevant to the question. However, it's unclear if these learned attention mechanisms truly reflect how humans attend to images to answer questions. This paper addresses the gap by empirically studying human attention in the context of VQA and comparing it directly to the attention generated by deep learning models. Understanding this alignment is crucial for developing more human-like and interpretable VQA systems.

Methods

The authors designed and implemented novel, game-inspired interactive interfaces for collecting human attention data. Subjects were presented with a blurred image and a question, and their task was to "sharpen" regions of the image using a mouse-based tool to reveal the information needed to answer the question. This "deblurring" process effectively creates an attention map where the sharpened areas indicate where the human focused. They experimented with different interface variations (e.g., showing the answer, showing the original image) to find the most effective way to collect informative attention maps.

To evaluate the VQA models, they used rank-order correlation as a metric. This involves scaling both human and machine-generated attention maps to a common resolution (14x14), ranking pixels by their attention weights, and then computing the correlation between these rankings. This method is robust to variations in the "peakiness" or "diffuseness" of attention maps. They also investigated the impact of "center bias" in attention maps, a known phenomenon where attention tends to be higher in the center of an image, by performing analysis on a subset of data with reduced center bias.

Datasets

The primary contribution in terms of datasets is the VQA-HAT (Human ATtention) dataset. This dataset was collected by the authors using their novel interfaces on a subset of the large-scale VQA dataset. They collected human attention maps for 58,475 training and 13,74 validation question-image pairs, involving approximately 20,000 Human Intelligence Tasks (HITs) performed by 800 unique workers. This dataset is intended to be released publicly.

The VQA dataset itself (Antol et al., 2015) is used as the source for the images and questions.

Key Insights and Ideas

  • Human attention is task-driven: Unlike general saliency prediction, human attention in VQA is highly influenced by the specific question being asked, focusing on relevant details (e.g., the floor for "What type is the surface?", player and racket for "Which game is being played?").
  • Current VQA attention models don't align with humans: The core finding is that attention maps generated by state-of-the-art VQA models show a surprisingly low correlation with human attention maps. This suggests that while these models might be accurate, their internal "focus" mechanisms are not mirroring human visual strategies.
  • Task-independent saliency is a better (but still imperfect) proxy: Interestingly, generic saliency models (like Judd et al.) showed a higher correlation with human VQA attention than VQA-specific attention models, even before accounting for center bias. This implies that general visual salience might be a more dominant factor in current VQA attention mechanisms than task-specific relevance.
  • Center bias is a factor, but not the whole story: While removing center bias improved the correlation for some models, the VQA-specific models still lagged behind human agreement, reinforcing the idea that their attention strategies are fundamentally different.
  • The VQA-HAT dataset is a valuable resource: The newly created dataset provides a benchmark for evaluating the human-likeness of VQA attention mechanisms and can be used for training attention-supervised VQA models.

Key Results

  • Low Rank Correlation: The most accurate VQA model evaluated (HieCoAtt-Q) achieved a mean rank-correlation of 0.26 with human attention maps. This is significantly lower than the inter-human agreement, which was 0.623.
  • Saliency vs. VQA Attention: Task-independent saliency maps (Judd et al.) had a mean rank-correlation of 0.497, which is higher than the VQA-specific models.
  • Impact of Center Bias: After controlling for center bias, the correlation for Judd saliency maps dropped significantly, while VQA-specific models showed a slight improvement in relative trends. HieCoAtt-Q achieved a higher correlation than Judd saliency on this reduced set, suggesting VQA-specific models can learn better task-relevant attention, but still not at human levels.
  • Human Accuracy: The "Blurred Image with Answer" interface yielded the highest human accuracy (78.7%) when subjects answered questions based on the sharpened images, indicating the quality of the collected attention maps.

Limitations and Future Work

  • Defining "Necessary" and "Sufficient" Attention: The authors acknowledge that defining what constitutes a "necessary" or "sufficient" attention map in the pixel space is an ill-posed problem. They pose this as an interesting question for future work: what is the right semantic space to discuss these concepts for humans?
  • Interpreting VQA Attention: The paper highlights that current VQA attention models are often "unsupervised" in the sense that attention is an intermediate variable optimized for downstream loss. The interpretability of these attention maps is often serendipitous. Future work could explore explicit supervision or different architectural choices to encourage more human-aligned attention.
  • Improving VQA Models: The findings suggest that future VQA models should aim to better mimic human attention strategies to improve both accuracy and interpretability. The VQA-HAT dataset is provided precisely for this purpose.
  • Further Validation: The authors mention that their dataset will allow for more thorough validation of their observations as new attention-based VQA models are developed.

Summary updated

Abstract

We conduct large-scale studies on `human attention' in Visual Question Answering (VQA) to understand where humans choose to look to answer questions about images. We design and test multiple game-inspired novel attention-annotation interfaces that require the subject to sharpen regions of a blurred image to answer a question. Thus, we introduce the VQA-HAT (Human ATtention) dataset. We evaluate attention maps generated by state-of-the-art VQA models against human attention both qualitatively (via visualizations) and quantitatively (via rank-order correlation). Overall, our experiments show that current attention models in VQA do not seem to be looking at the same regions as humans.
BibTeX
@article{das2016human,
  author = {Abhishek Das and Harsh Agrawal and C. Lawrence Zitnick and Devi Parikh and Dhruv Batra},
  title = {{Human Attention in Visual Question Answering: Do Humans and Deep Networks Look at the Same Regions?}},
  journal = {Comput. Vis. Image Underst.},
  year = {2016},
  url = {https://openreview.net/forum?id=Py7uOkJ9Q9},
}
Figure