Human-Adversarial Visual Question Answering

1 Introduction

• VQA v2 performance is approaching inter-human agreement, raising questions about how much of the VQA problem remains unsolved.The paper argues that AdVQA can illuminate current model shortcomings and that considerable room for improvement remains.Source: §1 · pp. 1–2
• Human annotators and state-of-the-art models interact in a dynamic collection process that uses model-fooling examples to stress test VQA systems.Other humans verify the examples; the collected data can serve as a subsequent benchmark iteration.Source: §1 · p. 1
• AdVQA contains 28,522 human-adversarial examples, all of which fooled the VQA 2020 challenge winner MoViE+MCAN.The paper describes this as the first application of the human-adversarial approach to a multimodal problem.Source: §1 · p. 2
• A wide range of existing VQA models perform significantly worse on AdVQA than on VQA v2 [18].The paper also analyzes AdVQA characteristics and contrasts them with VQA v2.Source: §1 · p. 2

2 Related Work

• AdVQA follows VQA stress tests of image grounding and specialized capabilities, while VQA v2’s near-human model performance contrasts with remaining unsaturated progress on the overall task.Prior datasets also test changing answer priors, text reading, external knowledge, spatial reasoning, and compositional reasoning.Source: §2 · pp. 2–3

2.1 Question Collection

• During question collection, annotators ask image-based questions while viewing a state-of-the-art model’s answer, and can submit after fooling it or trying at least ten questions.They may skip an image after three non-fooling questions; MoViE+MCAN is used in the loop to challenge collection against a current state-of-the-art model.Source: §2.1 · pp. 3–4, §2 · p. 3

2.2 Question Validation

• Two validators judge whether the model’s answer is incorrect and the question is valid; a third breaks ties, and at least two must agree on validity.A valid question requires the image to be necessary and sufficient for answering.Source: §2.2 · p. 4

2.3 Answer Collection

• The final answer-collection stage gathers ten human answers per validated question and includes checks for unanswerability, ambiguity, and poor annotator quality.The described quality controls include occasional hand-crafted questions with known, unambiguous answers.Source: §2.3 · pp. 4–5

2.4 Human-Adversarial Annotation Statistics

• On a subset with 100% human agreement, the model’s accuracy was 0.08% on AdVQA’s val split and 0.1% on its test split.Annotators took about five tries and around four minutes on average to find a model-fooling example.Source: §2.4 · p. 5
• The validated error rate for human-adversarial VQA was higher than ANLI’s 9.52% overall, while the claimed rate was similar to text-based adversarial QA rates of 44.0% for RoBERTa and 47.1% for BERT [4].The paper interprets this as evidence that fooling models is easier for VQA than for NLI.Source: §2.4 · p. 5

3.1 Baselines and Methods

• The evaluation includes answer-prior and unimodal baselines, multimodal pretrained models, and M4C OCR models trained on VQA or TextVQA and STVQA.Source: §3.1 · pp. 5–6
• Multimodal models generally outperform unimodal models, while the unimodal text model performs better than the unimodal image model on both datasets.Source: §3.1 · pp. 5–6

3.2 Discussion

• Most multimodal models do not outperform BERT or the per-answer-type majority baseline on AdVQA, and Table 4 shows near-prior performance across categories.The majority answers are “no” for yes/no, “2” for numbers, and “unanswerable” for others; M4C leads on numbers and others, possibly due to text reading.Source: §3.2 · p. 6, Table 4 · p. 7
• M4C performs best overall, VisualBERT leads among models without text reading, and all evaluated models perform poorly compared with humans.The authors suggest text reading and reasoning may help explain M4C’s performance; they note VQA performance is much closer to human performance.Source: §3.2 · pp. 6–7
• The authors flag possible overfitting to the model instance in the collection loop, but report that all evaluated VQA models perform poorly on AdVQA.They also note that model selection used the AdVQA validation set, unlike the model in the loop.Source: §3.2 · p. 7
• Finetuning VisualBERT and VilBERT on AdVQA validation data raises test accuracy from 28.70 to 33.10 and from 27.36 to 32.85, respectively.The authors take this as evidence of a difference between the VQA and AdVQA distributions.Source: §3.2 · p. 7

3.3 Training Details

• Models use author-provided architectures and hyperparameters, are finetuned from pretrained weights where applicable, and are averaged over three seeds.Training excludes COCO 2017 validation images; models are selected using AdVQA validation, with runs using up to eight V100 GPUs and lasting at most two days.Source: §3.3 · p. 7

• AdVQA contains 28,522 questions, each with 10 answers, and its questions average 7.8 words versus 6.3 for VQA v2 and 7.2 for TextVQA.The paper attributes the longer questions to workers making questions more specific to fool the model-in-the-loop.Source: §4 · p. 7
• AdVQA questions often begin with “what” or “how” and commonly ask about counts or text such as brands, signs, and times.The most frequent questions also include text-related and counting questions, identified as common failures for current state-of-the-art VQA models.Source: §4 · pp. 7–8, Figure 5 · p. 8
• Only 66.6% of AdVQA validation answers occur more than twice, versus 94.8% in VQA v2, indicating a more diverse answer vocabulary.The answer-vocabulary comparison suggests AdVQA is more diverse and difficult than VQA v2, but less narrowly focused than TextVQA.Source: §4 · p. 8, Figure 4 · p. 8
• 50.9% of AdVQA answers do not occur in VQA v2 training, although 77.2% of validation questions are answerable with the original VQA v2 vocabulary.The authors conclude that high performance requires understanding and reasoning about rare concepts, rather than a larger answer vocabulary alone.Source: §4 · pp. 8–9
• AdVQA shifts from VQA v2’s more frequent yes/no questions toward more number questions.The authors connect this shift to prior observations about VQA models’ difficulties with counting and text reading.Source: §4 · p. 9, Table 5 · p. 9
• In AdVQA validation, 97.7% of questions have an answer supported by at least three annotators, compared with 93.4% in VQA v2.The paper also reports unanimous agreement across all 10 answers for 27.2% of AdVQA questions.Source: §4 · p. 9
• 15.4% of AdVQA validation questions are answerable from Rosetta OCR tokens at edit distance 1, suggesting scene-text reading is important.Source: §4 · p. 9

5 Conclusion, Limitations & Outlook

• AdVQA is a novel human-adversarial multimodal dataset, and the authors conclude that VQA remains far from solved.The paper describes AdVQA as designed to accelerate progress on VQA.Source: §5 · p. 9
• Current state-of-the-art models underperform on AdVQA amid a reasoning gap involving text reading, counting, dataset bias, external knowledge, rare concepts, and multimodal understanding.The analysis argues that limited answer vocabulary, language representation, lack of pretraining, and lack of adversarial training are unlikely explanations for the gap.Source: §5 · p. 9
• AdVQA is intended to serve as a dynamic visual-reasoning benchmark with substantial headroom for further progress.The authors hope it will help bridge the gap.Source: §5 · p. 10
• Future work could continue AdVQA by putting new state-of-the-art models in the loop as they emerge.Source: §5 · p. 10

Broader Impact

• The authors hope that analyzing state-of-the-art VQA models with a dynamic human-adversarial dataset can help make models more robust.Source: Broader Impact · p. 10
• The work is English-only, and VQA datasets contain biases in their image, question, and answer distributions.The paper notes that image-captioning biases are also relevant to VQA.Source: Broader Impact · p. 10
• Common VQA datasets do not represent visually impaired users' images and questions well, and current models are not accurate enough to be useful or safe in these contexts.The paper points to VizWiz as a dataset of questions from visually impaired users about images they take for day-to-day tasks.Source: Broader Impact · p. 10

A Training Details

• Except for UNITER and VILLA, models were trained using MMF, generally with AdamW, a 5e−5 initial learning rate, cosine scheduling, and 2,000 warmup steps.Training used distributed PyTorch on NVIDIA V100 GPUs.Source: A Training Details · p. 14
• Training configurations varied by model; for example, M4C used its TextVQA schedule and OCR data, while UniT used a pretrained initialization.Source: A Training Details · pp. 14–15

B Off-the-shelf models

• Off-the-shelf OSCAR and VinVL validation performance on AdVQA was more than 2x lower than on VQA v2, despite overlap between their training data and the validation set.The passage directs readers to Table B.1 for validation performance.Source: Table B.1 · p. 15, B Off-the-shelf models · p. 15

C Annotation details

• Annotators were restricted to qualifying US-based, non-mobile workers and could earn a bonus when other annotators verified their question as fooling and valid.Source: C Annotation details · p. 15

C.1 Qualification Phase

• Qualification required reviewing 11 examples and passing a quiz with more than 6 of 7 answers correct; failed annotators received feedback and a second try.Source: §C.1 · p. 15

C.2 Main Labeling Task

• Instructions for first-time annotators covered question collection and question validation; returning annotators could access them through a button.The appendix also identifies Mechanical Turk preview landing pages.Source: §C.2 · p. 15
• Figure B.1 is labeled as showing examples from Qualification Phase Stage 1.Source: Figure B.1 · p. 16

C.3 Answer collection task

• Each question received ten answers from ten different annotators through the answer-collection interface.The section identifies Figures C.1 and C.2 as the preview and main interface.Source: §C.3 · p. 16, Figure C.1 · p. 19, Figure C.2 · p. 19
• Annotators received explicit instructions to reduce ambiguity and provide short, relevant answers, with an unanswerable checkbox they were told to use sparingly.The instructions followed [18] and [52].Source: §C.3 · p. 16
• Random hand-crafted questions with known answers screened annotators, who were barred from the task after three failed tests.Figures B.2 and B.3 show the qualification quiz and main-task instructions; access required passing the quiz.Source: §C.3 · p. 16, Figure B.2 · p. 17, Figure B.3 · p. 17

D Random Samples

• Other examples show variation in object or quantity labels, including “kettle” versus “teapot” and reports of one or two pieces of meat.Source: D Random Samples · pp. 21–22
• The ten samples include both strong consensus and varied responses, such as ten annotators answering “no” about window sizes and mixed answers about room occupancy.Table D.1 is identified as random examples from AdVQA.Source: D Random Samples · pp. 20–21, Table D.1
• For the largest-window question, answers vary across middle-window descriptions, while the processed answer is “middle” with a count of three.Raw responses include “center,” “bottom middle,” and other middle-window descriptions.Source: D Random Samples · p. 22
• Some raw answers converge on a response that processing marks as unknown, as with “sanssouci” and “pub.”Both examples are labeled “unknown (not in vocab)” in the processed answers.Source: D Random Samples · pp. 21–22