Abstract

Visual events are a composition of temporal actions involving actors spatially interacting with objects. When developing computer vision models that can reason about compositional spatio-temporal events, we need benchmarks that can analyze progress and uncover shortcomings. Existing video question answering benchmarks are useful, but they often conflate multiple sources of error into one accuracy metric and have strong biases that models can exploit, making it difficult to pinpoint model weaknesses. We present Action Genome Question Answering (AGQA), a new benchmark for compositional spatio-temporal reasoning. AGQA contains 192M unbalanced question answer pairs for 9.6K videos. We also provide a balanced subset of 3.9M question answer pairs, 3 orders of magnitude larger than existing benchmarks, that minimizes bias by balancing the answer distributions and types of question structures. Although human evaluators marked 86.02% of our question-answer pairs as correct, the best model achieves only 47.74% accuracy. In addition, AGQA introduces multiple training/test splits to test for various reasoning abilities, including generalization to novel compositions, to indirect references, and to more compositional steps. Using AGQA, we evaluate modern visual reasoning systems, demonstrating that the best models barely perform better than non-visual baselines exploiting linguistic biases and that none of the existing models generalize to novel compositions unseen during training.

Bibtex

@inproceedings{GrundeMcLaughlin2021AGQA,
title={AGQA: A Benchmark for Compositional Spatio-Temporal Reasoning},
author={Grunde-McLaughlin, Madeleine and Krishna, Ranjay and Agrawala, Maneesh},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year={2021}
}

AGQA Benchmark with programs & scene graphs

[Updated July 11 2021] This data is updated with the programs and scene graphs. There are slightly fewer questions due to fewer indirect object references per video. However, training and testing distributions have remained consistent.

Download Balanced AGQA with 3.5M questions (3.4G).

Download Unbalanced AGQA with 181M questions (199G).

Download Small unbalanced AGQA with 3.0M questions (3.2G).

AGQA Benchmark

This version of the benchmark is showcased in the CVPR 2021 paper.

Download Balanced AGQA with 3.9M questions (1.4G).

Download Unbalanced AGQA with 192M questions (71G).

Download Small unbalanced AGQA with 3.0M questions (1.1G).

Paper

Access our paper here.

README and supporting data

Download README and supporting data.

Scene Graphs

Download Scene Graphs of 9,601 videos (1.4G).

Code

Code is available here.