Commonsense Reasoning in Computer Vision: Foundations, Recent Advancements, and Future Directions
Bahar Uddin Mahmud, Sumit Barua, Guan Yue Hong, Ajay Gupta, Hexu Liu
cs.AI
Quick verdict
A broad survey of commonsense in vision; its systematic claim is undercut by unverifiable numbers, broken citations, and PRISMA arithmetic errors.
Summary
This survey reviews how commonsense knowledge is integrated into computer vision, covering knowledge graphs, scene graphs, neuro-symbolic models, and commonsense-enhanced vision transformers across classification, detection, segmentation, VQA, and captioning. From 1,050 retrieved records, a PRISMA-style process yields 44 included studies (2018–2025). The authors contribute a five-stage pipeline abstraction (perception, grounding, knowledge retrieval, fusion/reasoning, validation), a periodization into static, dynamic-retrieval, and implicit-knowledge eras (2019–2025), and a qualitative comparison of five models (SENet, NSCL, LXMERT, ViLT, Graph R-CNN). The strongest material analyzes failure modes: bias amplification in retrieved knowledge, prior–evidence conflicts, and cascading failures illustrated with medical and autonomous-driving examples. However, the survey contains no quantitative synthesis: Table 5 carries an 'Accuracy (%)' header filled with High/Medium labels, and the mAP/IoU-based per-paper assessment promised in §3.2 never appears. The PRISMA arithmetic is inconsistent (160−117 yields 43, not the stated 44), the inclusion taxonomy contradicts the methodology's categories, and several citations are broken or suspect—raw keys such as '(Chollet2019)' and '(li2022oscarpp)', an uncited 'SENet: Semantic Enrichment Network', a 'DeepLogic' attributed to DeepMind with no reference entry, and DETR misattributed to Li et al. 2021. Useful as an accessible map of the field, but its 'systematic' methodology and quantitative claims need verification.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Commonsense-enhanced vision systems share a canonical pipeline — perception, grounding, knowledge retrieval, fusion/reasoning, validation — enabling cross-paradigm comparison and gap analysis (§2.1, §5.4 Table 4).
- The field underwent a paradigm shift from static knowledge bases (pre-2019) through dynamic retrieval (2019–2021) to implicit parametric knowledge in foundation models (2022–2025) (§2.2).
- Retrieval-augmented vision reasoning is a distinct architectural class with multimodal-specific failure modes such as hallucination amplification and context mismatch (§2.3, §6.1).
- Retrieval can amplify rather than mitigate bias, and commonsense priors conflicting with visual evidence cause cascading failures in safety-critical settings (§8.3, §8.6, Tables 6–7).
- A PRISMA-style selection over 2018–2025 literature yields 44 studies forming the survey's evidence base (§3, Table 1).
Strengths
Genuinely reusable organizing abstraction
The five-stage pipeline (perception → grounding → knowledge retrieval → fusion/reasoning → validation) in §5.4/Table 4 maps concrete techniques, tools (ConceptNet, ATOMIC, FAISS, BLIP-2, LLaVA), and metrics onto each of five vision tasks — the most practically useful artifact in the paper.
Analytical periodization, not just a paper list
§2.2's three-era framework (static knowledge pre-2019 → dynamic retrieval 2019–2021 → implicit knowledge 2022–2025, with CLIP, KRISP, BEiT-3, Flamingo as anchors) gives real structure for understanding trade-offs between transparency and efficiency.
Above-average failure-mode treatment
Honest hedging of qualitative ratings
Weaknesses
No quantitative synthesis despite promising one
Systematic-review rigor gaps
Citation integrity problems
Invalid formalization presented as mathematics
Editorial quality undermines trust
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a survey of commonsense reasoning in computer vision, spanning knowledge graphs/scene graphs, neuro-symbolic models, and commonsense-enhanced vision transformers across five tasks (classification, detection, segmentation, VQA, captioning). Three contributions are claimed (§2): a unified pipeline abstraction ('Table 3' per §2.1, actually Table 4), a periodization static → dynamic-retrieval → implicit knowledge (2019–2025), and the first systematic treatment of retrieval-augmented vision reasoning (§6.1). The evidence base is 44 studies from a PRISMA-style selection (Table 1). Central theses: the field has shifted from explicit knowledge bases to implicit parametric knowledge; retrieval can amplify bias and produce cascading failures; future work should target causality, lightweight reasoning, trustworthiness, and explainability (§9).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is this paper about?
It is a survey of methods that inject commonsense knowledge into computer vision systems, covering knowledge graphs, scene graphs, neuro-symbolic models, and commonsense-augmented vision transformers across five tasks: classification, object detection, segmentation, visual question answering, and image captioning. It proposes a five-stage pipeline abstraction and discusses challenges such as bias, hallucination, and scalability.
Does the paper propose a new model or run experiments?
No. It reviews existing work — 44 studies selected through a PRISMA-style screening of about 1,050 records — and qualitatively compares five representative models (SENet, NSCL, LXMERT, ViLT, Graph R-CNN) using High/Medium/Low ratings. No new model, code, or benchmark numbers are reported.
What are the survey's main conclusions?
The field moved from explicitly curated knowledge bases (before 2019) to retrieval-augmented systems (2019–2021) and then to foundation models that internalize commonsense implicitly during pretraining (2022–2025). Retrieval can amplify bias and trigger cascading failures when priors contradict visual evidence; key open problems are causal reasoning beyond observables, real-time efficiency, bias mitigation, and explainability.
How reliable is it as a reference?
Partly. The topical coverage, pipeline abstraction, and failure-mode discussion are useful orientation for newcomers, but several citations are broken or unverifiable (an uncited 'SENet', a 'DeepLogic' with no reference entry, DETR misattributed to Li et al. 2021), the PRISMA counts are internally inconsistent, and quantitative claims in Section 8 should be checked against the original sources before being repeated.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–14 are visible only through their captions in the extraction; visual content (e.g., the Fig. 2 comparison of VLM outputs) could not be assessed.
- Table 4 and Table 5 cell values were read from linearized text; row/column pairings follow the extracted order.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.