PaperlayerView exact revision on arXiv

Source-linked AI summary

Roses Are Red, Violets Are Blue... but Should Vqa Expect Them To?

Corentin Kervadec, Grigory Antipov, Moez Baccouche, Christian Wolf

arXiv:2006.05121v3cs.CV

TL;DR

Imbalanced VQA data can let models exploit dataset biases, while overall accuracy gives limited insight into their errors on rare concepts. GQA-OOD evaluates accuracy on rare and frequent answers in context-specific shifted splits. The study finds that VQA models and bias-reduction methods still struggle to address rare question-answer pairs and OOD generalization.

  • Problem

    VQA datasets are imbalanced, and overall accuracy does not reveal how errors distribute across frequent and infrequent concepts or whether correct answers reflect reasoning or bias exploitation.

    Source: §1 · p. 1

  • Method

    GQA-OOD reorganizes GQA into context-specific shifted validation and test sets and evaluates accuracy on rare, common, and all answers within question groups.

    Source: §1 · p. 2, §3 · pp. 3–4

  • Results

    VQA models struggle on infrequent question-answer pairs, and neither conventional models nor dedicated bias-reduction methods succeed in all aspects of the benchmark.

    Source: §4.2 · p. 6, Discussion and conclusions · p. 8

  • Takeaways & Limitations

    The study indicates that substantial room remains for improving methods that address bias in VQA.

    Source: §1 · p. 2, Discussion and conclusions · p. 8

    Because GQA questions are automatically generated, the benchmark has limited vocabulary and synthetic syntax, despite using natural images.

    Source: §3 · p. 4

Abstract

from arXiv · show

Models for Visual Question Answering (VQA) are notorious for their tendency to rely on dataset biases, as the large and unbalanced diversity of questions and concepts involved and tends to prevent models from learning to reason, leading them to perform educated guesses instead. In this paper, we claim that the standard evaluation metric, which consists in measuring the overall in-domain accuracy, is misleading. Since questions and concepts are unbalanced, this tends to favor models which exploit subtle training set statistics. Alternatively, naively introducing artificial distribution shifts between train and test splits is also not completely satisfying. First, the shifts do not reflect real-world tendencies, resulting in unsuitable models; second, since the shifts are handcrafted, trained models are specifically designed for this particular setting, and do not generalize to other configurations. We propose the GQA-OOD benchmark designed to overcome these concerns: we measure and compare accuracy over both rare and frequent question-answer pairs, and argue that the former is better suited to the evaluation of reasoning abilities, which we experimentally validate with models trained to more or less exploit biases. In a large-scale study involving 7 VQA models and 3 bias reduction techniques, we also experimentally demonstrate that these models fail to address questions involving infrequent concepts and provide recommendations for future directions of research.

1. Introduction

VQA models are expected to reason over images and questions, but imbalanced data can encourage shortcuts and make overall accuracy an inadequate diagnostic. The paper proposes a fine-grained GQA benchmark and evaluation protocol, then studies how current models and bias-reduction methods perform.

1. Introduction

  • Imbalanced VQA data can encourage models to rely on dataset biases, while overall accuracy leaves error distributions and performance on rare concepts unclear.The paper raises whether correct answers reflect reasoning or bias exploitation, and how models perform on infrequent versus frequent concepts.Source: §1 · p. 1
  • GQA-OOD reorganizes GQA with distribution shifts in validation and test, alongside metrics and plots for evaluating generalization across operating points.Its question groups support analysis of errors across concept associations and their frequencies.Source: §1 · p. 2
  • The benchmark is presented as complementary to existing protocols and as an alternative to VQA-CP, whose results have been mixed.The authors motivate GQA's question-group structure as a way to tailor shifts to each question's context.Source: §1 · pp. 1–2
  • A study of VQA models and bias-reduction methods finds that they struggle to generalize under OOD conditions, leaving room to improve bias reduction.The contributions include evaluating the proposed metric on models trained to exploit biases to different degrees.Source: §1 · p. 2

2. Related Work

Prior work documents dataset biases, attempts to reduce them, and the limits of existing evaluation protocols. These limitations motivate a protocol that tests both in- and out-of-distribution behavior without tailoring methods to a single handcrafted shift.

VQA corpuses

  • VQA1 had ~76K questions over 25K realistic images; proposed VQA2 to address strong answer imbalances, while synthetic CLEVR was adapted to real-world images as GQA (1.7M questions), offering better control on dataset statistics.Source: §2 · p. 2

Attempts to reduce the bias-dependency

  • VQA models can rely on question biases, overlook relevant visual regions, and fail to generalize to unseen distributions.Bias-reduction approaches include adversarial training, RUBi, question-type statistics, and human-guided visual attention.Source: §2 · p. 2

Reinventing VQA evaluation

  • Existing VQA metrics and splits offer partial diagnostics: GQA metrics do not test OOD correctness, while other benchmarks examine specific forms of generalization.VQA-CP2 changes the train-test answer distribution, and its evaluation has been criticized for test-set tuning and other protocol pitfalls.Source: §2 · pp. 2–3

Shortcomings of Out-of-distribution testing

  • Critiques of OOD evaluation include the lack of a validation split and methods tailored to known test-shift construction procedures.A study found random visual attention at least as effective as human-relevant attention on OOD and in-distribution settings.Source: §2 · p. 3
  • The paper proposes a protocol for testing both in- and out-of-distribution settings, paired with a study of VQA architectures and debiasing methods.The authors frame this as a way to better understand those methods' strengths and weaknesses.Source: §2 · p. 3

3. GQA-OOD: a benchmark for OOD settings

GQA-OOD creates context-specific distribution shifts using GQA question groups, then evaluates accuracy on rare and common answers. Its validation split also supports OOD hyperparameter selection, while its reliance on automatically generated questions limits vocabulary and syntax.

3. GQA-OOD: a benchmark for OOD settings

  • GQA-OOD defines OOD samples as rare events relative to a base distribution and makes shifts depend on question context.For “What color is this rose?”, blue is rare; for a violet, red may be rare instead.Source: §3 · p. 3
  • Table 1 reports data statistics for GQA-OOD versus GQA and for head versus tail samples.The caption identifies the table's comparison categories but does not state specific values.Source: Table 1 · p. 3
  • The benchmark keeps GQA's training set and constructs shifted validation and test subsets by grouping questions, selecting imbalanced groups, and choosing OOD samples.The shifted sets use the existing GQA training split.Source: §3 · p. 3

Question groups

  • GQA local groups capture question type and related concepts, structuring the benchmark around context-specific answer distributions.The validation split contains about 37K groups related to about 132K questions; smoothing the balanced GQA distribution does not remove real-world imbalances.Source: §3 · p. 3

Measuring group imbalance

  • Groups are selected for meaningful imbalance using normalized Shannon entropy, retaining those below an empirically set threshold of T=0.9.The normalization accounts for the varying number of possible answers across groups.Source: §3 · pp. 3–4

OOD setting and metrics

  • Acc-tail measures accuracy on rare answers, Acc-head on the remaining answers, and Acc-all on all benchmark samples.Tail classes have counts at or below α times the group's mean count, with α empirically set to 1.2.Source: §3 · p. 4

Difference with VQA-CP2

  • Compared with VQA-CP2, GQA-OOD supports question-group analysis, metrics for different shift amounts, and continuous evaluation across operating points.It also includes a shifted validation set for selecting hyperparameters under OOD conditions, while requiring training on the existing GQA split.Source: §3 · p. 4

Limitations

  • The benchmark's questions are automatically generated, limiting vocabulary and producing synthetic syntax, although its images are natural and real.GQA covers 70% of VQA2 answers.Source: §3 · p. 4

4. Experiments

The experiments assess Acc-tail as a diagnostic of bias sensitivity, then use GQA-OOD to compare model generalization, bias-reduction methods, and established benchmarks.

4.1. Evaluation of the proposed metric

  • 1.4% relative difference on GQA overall contrasts with BUTD+bal’s large Acc-tail advantage over BUTD+all, widening as α decreases.The two models are BUTD variants trained on balanced and raw, unbalanced GQA training sets, respectively.Source: Table 2 · p. 5, §4.1 · pp. 5–6

4.2. Analysis of VQA model error distributions

  • About 30 points: tail accuracy falls below overall accuracy as α narrows, while VIS-ORACLE stays nearly flat; LXMERT head/tail confusion rises from 25% to 42%.The LXMERT confusion comparison is for α=1.2 and α=0.3, respectively.Source: Figure 3 · p. 5, §4.2 · p. 6
  • Table 3’s Acc-tail comparison shows that VQA models struggle on rare pairs, though LXMERT leads; Acc-all is mechanically increased by high scores on frequent answers.Source: §4.2 · p. 6, Table 3 · p. 7
  • LXMERT’s proxy labels suggest reasoning on verify, choose, and logical questions but bias on query and compare; only 1% of tail questions are compare.Source: §4.2 · p. 7, Figure 5 · p. 7, §4.3 · p. 7

4.3. Re-evaluating bias-reduction methods

  • None of RUBi, BP, or LM improves Acc-tail; frequent-end accuracy falls while rare-end accuracy stays nearly unchanged, so bias reduction need not aid rare-pair generalization.The curve comparison is shown for bias-reduction methods applied to BUTD.Source: Figure 4 · p. 6, §4.3 · pp. 7–8

4.4. Comparison with other benchmarks

  • About 10 points: LXMERT loses this much in GQA-OOD; despite matching LXMERT on GQA overall, MMN struggles more on rare question-answer pairs.The comparison also finds overall accuracy on GQA and VQA2 insufficient to fully characterize VQA performance.Source: Table 4 · p. 8, §4.4 · p. 8
  • 52.0% VQA-CP2 accuracy for LM coexists with 33% GQA-OOD Acc-tail, so VQA-CP2 performance does not reflect behavior on rare question-answer pairs.Source: Table 4 · p. 8, §4.4 · p. 8

Discussion and conclusions

GQA-OOD evaluates bias absorption and generalization to rare or unseen question-answer pairs. The experiments show that neither conventional state-of-the-art VQA models nor dedicated bias-reduction methods succeed across all benchmark aspects.

Discussion and conclusions

  • GQA-OOD combines imbalanced question groups, new metrics, and adjustable distribution shift through α to evaluate bias absorption and generalization to rare or unseen pairs.The benchmark is intended to assess both whether models absorb training-data tendencies and how well they generalize.Source: Discussion and conclusions · p. 8
  • Experiments show that neither conventional state-of-the-art VQA models nor dedicated bias-reduction methods succeed in all aspects of the benchmark.Source: Discussion and conclusions · p. 8
  • The authors hope GQA-OOD will encourage models less prone to spurious biases and more reliable in real-world scenarios.Source: Discussion and conclusions · p. 8

A. Additional examples from the GQA-OOD validation split

Validation examples illustrate the imbalance of tail answer groups and varied model predictions. Across these examples, models mostly rely on context biases, while most models answer the shown head question correctly.

A. Additional examples from the GQA-OOD validation split

B. Dataset statistics

The benchmark analysis examines question distributions and reports that GQA-OOD preserves GQA’s question diversity. Its validation split has a larger proportion of open questions, which generally allow more varied answers and are more prone to imbalance.

B. Dataset statistics

  • The analysis assesses GQA-OOD question distributions and reports that the benchmark preserves GQA’s original question diversity.Source: B. Dataset statistics · p. 11
  • GQA-OOD preserves the original split’s distribution of question structure types, while the proportion of open “query” questions increases.The cited figures compare validation-split question structures in GQA and GQA-OOD.Source: B. Dataset statistics · p. 11
  • Open questions such as color questions allow a wider diversity of answers and are therefore more prone to imbalance than “choose,” “verify,” or “compare” questions.The latter types usually accept only two possible answers and are easier to balance.Source: B. Dataset statistics · p. 11

C. Training details

The appendix reports model-training settings and implementation details, along with analyses of rare-answer predictions and the LM entropy penalty. The reported penalty degrades GQA-OOD accuracy when training on GQA.

C. Training details

  • Most listed models were trained for 20 epochs with batch size 512 and Adam, while MCAN used 11 epochs, batch size 64, and Adamax.LSTM, BUTD, RUBi, BP, and LM used the first setup; MCAN used the second.Source: C. Training details · p. 11
  • For most models, rarer ground-truth answers make predictions more likely to belong to the head set.The analysis progressively reduces the tail set while keeping the head set unchanged.Source: C. Training details · p. 12
  • LXMERT pretraining samples from the GQA validation split were removed before pretraining to support evaluation on GQA and GQA-OOD validation splits.The pretraining corpus contained images and sentences from MSCOCO and VisualGenome.Source: C. Training details · p. 12
  • The VIS-ORACLE uses a small LXMERT architecture and ground-truth GQA objects encoded with class, attribute, and scene-graph relationship one-hot vectors.Source: C. Training details · p. 12
  • The LM entropy penalty degrades GQA-OOD accuracy when training on GQA, with higher penalty weights flattening the curve for frequent samples further.Source: C. Training details · p. 12, Figure 10 · p. 12

D. Measuring head/tail confusion

The section measures head/tail confusion to assess how far prediction errors relate to context-bias dependence. Validation examples contrast model predictions on tail and head questions.

D. Measuring head/tail confusion

  • Head/tail confusion is measured as a proportion of questions to examine whether prediction errors relate to context-bias dependence.The main paper omits the procedure's detailed description and refers readers to a following paragraph for the exact methodology.Source: D. Measuring head/tail confusion · p. 12
  • On three tail examples, model predictions vary, with several models missing the answers “bridge,” “brown,” and “glove.”The accompanying histograms represent answer frequency across all questions in each question group.Source: Figure 11 · p. 13, Figure 12 · p. 13, Figure 13 · p. 14
  • On the head example, all evaluated models except LSTM predict the answer “dog.”Its histogram represents answer frequency across all questions in the question group.Source: Figure 14 · p. 14
Loading 2006.05121v3…