VQA
收藏资源简介:
我们提出了自由形式和开放式视觉问答 (VQA) 的任务。给定图像和关于图像的自然语言问题,任务是提供准确的自然语言答案。反映许多现实世界的场景,例如帮助视障人士,问题和答案都是开放式的。视觉问题有选择地针对图像的不同区域,包括背景细节和底层上下文。因此,与生成通用图像说明的系统相比,在 VQA 上取得成功的系统通常需要对图像和复杂推理有更详细的理解。此外,VQA 适合自动评估,因为许多开放式答案仅包含几个单词或一组封闭的答案,可以以多项选择的形式提供。我们提供了一个数据集包含 100,000 的图像和问题并讨论它提供的信息。提供了许多 VQA 基线,并与人类表现进行了比较。
We propose the task of free-form and open-ended visual question answering (VQA). Given an image and a natural language question regarding the image, the task requires generating an accurate natural language answer. Reflecting numerous real-world scenarios such as assisting visually impaired individuals, both the questions and answers are open-ended. Visual questions selectively target different regions of the image, including background details and underlying context. Therefore, compared to systems that generate generic image captions, successful VQA systems typically demand more detailed comprehension of images and complex reasoning capabilities. Moreover, VQA is amenable to automatic evaluation, as many open-ended answers only consist of a few words, or a set of closed-ended answers can be provided in the form of multiple-choice questions. We present a dataset containing 100,000 images and questions, and discuss the information it encompasses. Multiple VQA baselines are provided and compared against human performance.

- 首次提出视觉问答(VQA)任务,并发布了第一个VQA数据集,包含204,721个图像和82,783个问题。
- 发布了VQA v1.0数据集,包含200,000个问题和400,000个答案,成为视觉问答领域的基准数据集。
- 发布了VQA v2.0数据集,包含110万对问题和答案,旨在解决v1.0中存在的偏差问题,提升数据集的挑战性和公平性。
- VQA数据集被广泛应用于多个视觉问答挑战赛,推动了相关算法和模型的快速发展。
- 发布了VQA-CP v2数据集,旨在通过改变问题和答案的配对方式,进一步提高模型的泛化能力。
- 1VQA: Visual Question AnsweringUniversity of Toronto, University of Montreal · 2015年
- 2Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question AnsweringUniversity of Michigan, University of Toronto · 2017年
- 3Bottom-Up and Top-Down Attention for Image Captioning and Visual Question AnsweringUniversity of Adelaide, Microsoft Research · 2018年
- 4GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question AnsweringUniversity of California, Berkeley · 2019年
- 5Counterfactual Samples Synthesizing for Robust Visual Question AnsweringTsinghua University, University of California, Los Angeles · 2020年



