Integrating Domain Knowledge into Large Language Models for Enhanced Fashion Recommendations

Shanglin Yang; Zhan Shi

REVIEW 4 major objections 4 minor 1 cited by

Integrating Domain Knowledge into Large Language Models for Enhanced Fashion Recommendations

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper's Fashion Large Language Model, fine-tuned with auto-prompt generation and retrieval-augmented inference, achieves the highest fill-in-the-blank accuracy on Polyvore among compared methods.

desk verdict A plausible LLM-for-fashion pipeline whose accuracy claims are undercut by a modality confound and a potentially leaking retrieval step. read the letter →

arxiv 2502.15696 v1 pith:SAMCG2TZ submitted 2025-01-03 cs.CL cs.AIcs.IRcs.LG

Zhan Shi , Shanglin Yang This is my paper · ORCID

classification cs.CLcs.AIcs.IRcs.LG

keywords fashionrecommendationlargelanguagemodelsretrieval-augmentedgenerationfill-in-the-blankPolyvorefew-shotlearningauto-promptoutfitcompatibility

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes Fashion Large Language Model (FLLM), an LLM fine-tuned on outfit text data with auto-prompt generation, plus retrieval-augmented inference, for fashion recommendation. It argues that this text-based generative approach outperforms established image-based compatibility models on the fill-in-the-blank task: FLLM reaches 62.17 on Polyvore-disjoint and 67.21 on Polyvore, the highest in the comparison table. The paper also claims that FLLM retains accuracy at low training-data ratios, surpassing a Type-Aware baseline by about 10 percent when data is scarce. If these results hold, they suggest that language models can substitute for visual models in fashion compatibility reasoning while adding interpretability and adaptability.

What carries the argument

The central machinery is two-fold: (1) an auto-prompt generation training strategy in which both template QA pairs and LLM-generated, style-focused QA pairs are used to fine-tune the model on Polyvore outfits, preserving domain knowledge; and (2) a retrieval-augmented generation (RAG) inference stage that issues multiple query pathways — direct embedding queries, style- and occasion-conditioned queries, and LLM-generated dynamic questions — against a vector database of fashion documents. Together these components let the model condition its recommendations on both broad fashion knowledge and a user's stated preference context.

What would settle it

A decisive test would rerun the fill-in-the-blank evaluation on Polyvore-disjoint with a simple text-only baseline (for example, a bag-of-words or BERT classifier over item names and category labels) and with an FLLM variant whose inputs are stripped of category and attribute words; if either matches or exceeds FLLM's 62.17, the claimed advantage would be attributable to reading labels rather than to LLM-based reasoning or retrieval.

Watch

Extended reading notes

Core claim

The central claim is that a fine-tuned LLM with retrieval augmentation can solve outfit fill-in-the-blank prediction more accurately than specialized image-based models. On the Polyvore-disjoint test set, FashionLLM achieves 62.17 percent accuracy, ahead of OutfitTransformer at 59.48, CSA-Net at 59.26, Type-Aware at 55.65, and SCE-Net at 53.67; on the full Polyvore set it reaches 67.21, edging OutfitTransformer's 67.10. The paper also reports that at low training-data ratios FLLM beats the Type-Aware baseline by roughly 10 percentage points, and that its generated recommendations are interpretable and versatile across style preferences.

Load-bearing premise

The load-bearing premise is that the Table 1 comparison is fair, meaning FLLM receives item text while the baselines receive only images and the accuracy gap reflects better fashion reasoning rather than the model reading category or attribute labels from the text, a premise never tested with a text-only baseline.

Editorial extensions

If this is right

If FLLM's accuracy advantage holds, outfit compatibility can be treated as a language task, enabling recommendation systems to leverage pretrained reasoning and user-facing explanation in one model.
The reported few-shot robustness implies that text-based LLM recommenders can be deployed in data-sparse settings such as new boutiques, niche styles, or emerging trends where visual training data is scarce.
The RAG-based inference with multiple query pathways suggests a modular design: updating the vector database can refresh the model's fashion knowledge without retraining the LLM itself.
The interpretability of generated outfit suggestions could support direct user interaction, allowing users to state preferences and receive style-alternative recommendations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The comparison in Table 1 is between a text-fed LLM and image-fed baselines; if Polyvore item descriptions encode category or attribute labels, part of the observed gain may come from the model reading labels the baselines must infer from pixels. A text-only baseline or a description-stripped ablation would isolate the language-model reasoning contribution.
Because the fine-tuning uses Polyvore training outfits and the RAG database stores fashion documents, the model's 'domain knowledge' may be partly memorized from the dataset; testing on a held-out trend shift (e.g., newer outfits) would show whether retrieval actually helps adapt to distribution shift as claimed.
The few-shot claim centers on one baseline (Type-Aware); extending the low-data comparison to other image-based methods and reporting variance across seeds would strengthen the generalizability of the 10 percent figure.
A natural next step is a multimodal hybrid where the LLM fuses its text reasoning with image encoders, which the authors themselves signal in their future-work section; such a hybrid could disambiguate whether text or image contributes the compatibility signal.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a plausible LLM-for-fashion pipeline, but the paper's headline accuracy claims rest on a comparison that is probably unfair, and the retrieval step may leak the answer. The core idea—fine-tune a small LLM on auto-generated QA and add RAG at inference—is legit and cheap, and the FITB numbers on Polyvore are at least reported on a standard benchmark. Credit where due: the authors clearly thought about training-data construction and they went to the trouble of evaluating on Polyvore-disjoint, which is more than many LLM papers do.

The soft spots are real and they hit the central claim. In Table 1, FLLM consumes item text while Type-Aware, CSA-Net, and OutfitTransformer consume images. If that text carries category or attribute names, a language prior can solve category-level FITB without fashion reasoning. There is no text-only baseline, no category-masking experiment, no error bars. Second, the retrieval-augmented inference in Section 3.3 embeds item descriptions and pulls them from a vector DB. Nothing in the paper says the candidate items are excluded from that DB. If they are in, the model can just copy the retrieved description. A RAG on/off ablation would settle this; none is reported. The few-shot figure also lacks axis labels and variance, and the 'interpretability advantage' is asserted but never measured.

The paper also claims there is 'no established work on using LLMs in fashion recommendation,' which is not credible—there is prior work on LLMs for outfit generation and compatibility, and the citation list doesn't engage with it.

None of this makes the method incoherent. The approach could be made to work; it's just that the evidence as presented does not support the superiority claim. As is, this reads like a workshop submission that needed one more round of honest ablation work before being put up for serious review. I would not send it to peer review in this state, but I'd tell the authors to redo the evaluation with a text-only baseline, a RAG-ablation, and variance bars; then it might become a solid domain application.

Referee Report

4 major / 4 minor

Summary. The paper proposes FLLM, a large language model fine-tuned on fashion-specific QA data derived from Polyvore and equipped with retrieval-augmented generation at inference. The authors report fill-in-the-blank (FITB) accuracy on Polyvore and Polyvore-disjoint, few-shot learning results at reduced training-data ratios, and a qualitative visualization of recommendation diversity. The main claims are that FLLM achieves the highest FITB accuracy in Table 1, surpasses a Type-Aware baseline by about 10 percent in few-shot settings, and offers improved interpretability.

Significance. If the reported results were valid, the paper would make a useful contribution: a text-only LLM with retrieval augmentation outperforming established image-based compatibility models on a public benchmark would be a noteworthy result with practical implications for explainable fashion recommendation. The use of a fixed external benchmark, Polyvore, is a strength: accuracy is measured against ground-truth labels rather than the model's own outputs, so the central accuracy claim is not circular. The paper also provides a clear, if high-level, architecture for combining auto-prompt fine-tuning with RAG. However, the experimental design has load-bearing confounds that prevent the claims from being accepted as stated.

major comments (4)

[Section 4.2, Table 1] The main comparison is confounded by input modality. FLLM consumes item text while Type-Aware, CSA-Net, and OutfitTransformer consume images; if the item descriptions include category or attribute labels, the FITB task can be solved partially or entirely by applying a language prior to those labels. No text-only baseline, category-masking experiment, or input-modality ablation is provided, so the reported accuracy advantage in Table 1 cannot be attributed to fashion compatibility reasoning rather than to the information available in the text modality.
[Section 3.3, Table 1] The retrieval-augmented inference setup creates a possible leakage channel for the FITB evaluation. The vector database described in Section 3.3 is not stated to exclude the candidate items that appear in the FITB test questions, and the paper reports no RAG on/off ablation, no retrieval hit-rate analysis, and no check for whether the retrieved item text contains the answer. If the database includes candidate item descriptions, the model can achieve high FITB accuracy simply by adopting the top retrieved item's text, which would make the numerical results in Table 1 uninformative about fashion reasoning.
[Section 4.2, Table 1 and Figure 4] No error bars, standard deviations, number of independent runs, or significance tests are reported for any of the accuracy numbers. The few-shot claim of 'about 10 percent' improvement appears to rest on single runs at low data ratios, and the statement that final performance is similar at the full data ratio is not enough to establish a reliable difference. Additionally, the axes of Figure 4 are internally inconsistent: the x-axis lists both 0%-80% and 10%-50%, and the y-axis is labeled 'FIIB accuracy' while the text uses FITB, making the figure difficult to interpret.
[Section 4.2, Figure 5] The claimed interpretability advantage is supported only by a single qualitative visualization. There is no user study, human evaluation, or quantitative faithfulness metric for the generated explanations or recommendations, so the statement that FLLM 'surpasses existing models in accuracy, interpretability, and few-shot learning capabilities' is not demonstrated for the interpretability component.

minor comments (4)

[Table 1] The SCE-Net Average row cites reference [22], which is a remote sensing paper on single-view height estimation and semantic segmentation, not a fashion compatibility method; the provenance of this baseline is therefore unclear and needs to be corrected or replaced.
[Section 3.2] The fine-tuning procedure omits essential reproducibility details: the base LLM is not named, and the number of training steps, learning rate, data sizes, and other hyperparameters are not reported. Section 3.3 similarly omits the sentence-transformer model and the retrieval top-k setting.
[Section 4.2] The text uses 'FITB' while Figure 4 uses 'FIIB' for the same task; this inconsistency should be fixed, and the figure legend should clarify which curves correspond to which method.
[References] The manuscript retains template placeholders in the ACM reference format ('https://doi.org/XXXXXXX.XXXXXXX') and in the acknowledgments section, which should be completed before any publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FLLM's accuracy claims are measured against the external Polyvore benchmark, and no derivation step reduces to its own inputs by construction.

full rationale

The paper's central claims are empirical accuracy comparisons on Polyvore Outfits and Polyvore Outfits-D (Table 1) and few-shot accuracy ratios (Figure 4). These are evaluated against a fixed, externally published benchmark, so the evaluation is not defined in terms of FLLM's own outputs. The method sections describe template QA generation and LLM auto-QA generation (Section 3.2); while the LLM generates some training prompts, this is a data-augmentation step and does not make the held-out accuracy measure equal to the training objective by construction. The RAG module (Section 3.3) retrieves fashion documents to condition inference; potential leakage of candidate items into the retrieval index would threaten validity, but the paper does not state that the test candidates are in the index, and absent a quoted reduction this remains a speculation rather than a demonstrated circular step. No load-bearing self-citations appear; references to the authors' own prior work are absent, and no fitted parameter is renamed as a prediction. Therefore, no step reduces to its own inputs by definition.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper offers no derivation or closed-form result. Its central claim rests on domain assumptions about the Polyvore benchmark being a valid test of fashion compatibility, about text descriptions being comparable to images, and about self-generated prompts improving the model. The only hand-chosen inputs are unreported training and retrieval hyperparameters.

free parameters (1)

Unreported training and retrieval hyperparameters
The paper does not report learning rate, number of epochs, LoRA rank, retrieval top-k, or the sentence transformer model, so any reproduction requires guessing these hand-chosen values.

assumptions (3)

domain assumption Polyvore outfit compatibility labels are a reliable ground truth for fashion compatibility.
Used as the benchmark for FITB accuracy in Section 4.2 without discussion of label noise or subjectivity.
domain assumption Textual item descriptions are sufficient for the FITB task, so comparisons with image-based models are valid.
The LLM is fed text, while baselines use images; the paper does not justify modality equivalence.
domain assumption Auto-prompt generated QA pairs improve, rather than distort, domain knowledge.
The method in Section 3.2 assumes LLM-generated fashion questions and answers provide useful supervision, with no validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Integrating Domain Knowledge into Large Language Models for Enhanced Fashion Recommendations." pith.science (2026). https://pith.science/paper/SAMCG2TZ

@misc{pith2026250215696,
  author       = {Pith},
  title        = {Pith review of: Integrating Domain Knowledge into Large Language Models for Enhanced Fashion Recommendations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SAMCG2TZ}},
  note         = {Machine review of arXiv:2502.15696}
}

read the original abstract

Fashion, deeply rooted in sociocultural dynamics, evolves as individuals emulate styles popularized by influencers and iconic figures. In the quest to replicate such refined tastes using artificial intelligence, traditional fashion ensemble methods have primarily used supervised learning to imitate the decisions of style icons, which falter when faced with distribution shifts, leading to style replication discrepancies triggered by slight variations in input. Meanwhile, large language models (LLMs) have become prominent across various sectors, recognized for their user-friendly interfaces, strong conversational skills, and advanced reasoning capabilities. To address these challenges, we introduce the Fashion Large Language Model (FLLM), which employs auto-prompt generation training strategies to enhance its capacity for delivering personalized fashion advice while retaining essential domain knowledge. Additionally, by integrating a retrieval augmentation technique during inference, the model can better adjust to individual preferences. Our results show that this approach surpasses existing models in accuracy, interpretability, and few-shot learning capabilities.

Figures

Figures reproduced from arXiv: 2502.15696 by the authors.

**Figure 2.** An overview of an information-augmented retrieval system featuring a fine-tuned Large Language Model (Fashion [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Detailed Workflow of Model Fine-Tuning and Re [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Performance of different methods with respect to [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Visualization of recommendation results: For each [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

FashionM3: Multimodal, Multitask, and Multiround Fashion Assistant based on Unified Vision-Language Model
cs.CV 2025-04 conditional novelty 6.0 of 10

A fine-tuned vision-language model trained on a new FashionRec dataset outperforms off-the-shelf VLMs on held-out dialogue similarity metrics for fashion recommendation.

Reference graph

Works this paper leans on

25 extracted references · 15 canonical work pages · cited by 1 Pith paper

[1]

Ziad Al-Halah, Rainer Stiefelhagen, and Kristen Grauman. 2017. Fashion for- ward: Forecasting visual style in fashion. In Proceedings of the IEEE international conference on computer vision . 388–397

work page 2017
[2]

Victor Carbune, Hassan Mansoor, Fangyu Liu, Rahul Aralikatte, Gilles Baechler, Jindong Chen, and Abhanshu Sharma. 2024. Chart-based reasoning: Transferring capabilities from llms to vlms. arXiv preprint arXiv:2403.12596 (2024)

arXiv 2024
[3]

Wen-Huang Cheng, Sijie Song, Chieh-Yun Chen, Shintami Chusnul Hidayati, and Jiaying Liu. 2021. Fashion Meets Computer Vision: A Survey. ACM Comput. Surv. 54, 4, Article 72 (jul 2021), 41 pages. https://doi.org/10.1145/3447239

work page doi:10.1145/3447239 2021
[4]

Yujuan Ding, Zhihui Lai, PY Mok, and Tat-Seng Chua. 2023. Computational Technologies for Fashion Recommendation: A Survey. Comput. Surveys 56, 5 (2023), 1–45

2023
[5]

Xintong Han, Zuxuan Wu, Yu-Gang Jiang, and Larry S Davis. 2017. Learning fashion compatibility with bidirectional lstms. In Proceedings of the 25th ACM international conference on Multimedia . 1078–1086

work page 2017
[6]

Hyunwoo Hwangbo, Yang Sok Kim, and Kyung Jin Cha. 2018. Recommendation system development for fashion retail e-commerce.Electronic Commerce Research and Applications 28 (2018), 94–101

2018
[7]

Pei Ke, Bosi Wen, Zhuoer Feng, Xiao Liu, Xuanyu Lei, Jiale Cheng, Shengyuan Wang, Aohan Zeng, Yuxiao Dong, Hongning Wang, Jie Tang, and Minlie Huang

work page
[8]

Yen-Liang Lin, Son Tran, and Larry S. Davis. 2020. Fashion Outfit Complementary Item Retrieval. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE. https://doi.org/10.1109/cvpr42600.2020.00337

arXiv 2020

Show all 25 references

[9]

Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2024. Once: Boosting content-based recommendation with both open-and closed-source large language models. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining. 452–461

2024
[10]

Xin Liu, Yongbin Sun, Ziwei Liu, and Dahua Lin. 2020. Learning diverse fashion collocation by neural graph filtering. IEEE Transactions on Multimedia 23 (2020), 2894–2901

2020
[11]

Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, and Jiebo Luo. 2023. Llm- rec: Personalized recommendation via prompting large language models. arXiv preprint arXiv:2307.15780 (2023)

2023 arXiv
[12]

Dongmei Mo, Xingxing Zou, Kaicheng Pang, and Wai Keung Wong. 2023. To- wards private stylists via personalized compatibility learning. Expert Systems with Applications 219 (2023), 119632

2023
[13]

Stefanos-Iordanis Papadopoulos, Christos Koutlis, Symeon Papadopoulos, and Ioannis Kompatsiaris. 2022. Multimodal Quasi-AutoRegression: Forecasting the visual popularity of new fashion products. International Journal of Multimedia Information Retrieval 11, 4 (2022), 717–729

2022
[14]

Stefanos-Iordanis Papadopoulos, Christos Koutlis, Symeon Papadopoulos, and Ioannis Kompatsiaris. 2022. VICTOR: Visual Incompatibility Detection with Transformers and Fashion-specific contrastive pre-training. arXiv preprint arXiv:2207.13458 (2022)

2022 arXiv
[15]

Rohan Sarkar, Navaneeth Bodla, Mariya I Vasileva, Yen-Liang Lin, Anurag Beni- wal, Alan Lu, and Gerard Medioni. 2023. Outfittransformer: Learning outfit representations for fashion recommendation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vis...

2023
[16]

Maria Anastassia Stefani, Vassilios Stefanis, and John Garofalakis. 2019. CFRS: a trends-driven collaborative fashion recommendation system. In 2019 10th Inter- national Conference on Information, Intelligence, Systems and Applications (IISA) . IEEE, 1–4

2019
[17]

Vasileva, Bryan A

Mariya I. Vasileva, Bryan A. Plummer, Krishna Dusad, Shreya Rajpal, Ranjitha Kumar, and David Forsyth. 2018. Learning Type-Aware Embeddings for Fashion Compatibility. In Proceedings of the European Conference on Computer Vision (ECCV)

2018
[18]

Mariya I Vasileva, Bryan A Plummer, Krishna Dusad, Shreya Rajpal, Ranjitha Kumar, and David Forsyth. 2018. Learning type-aware embeddings for fashion compatibility. InProceedings of the European conference on computer vision (ECCV). 390–405

2018
[19]

Andreas Veit, Balazs Kovacs, Sean Bell, Julian McAuley, Kavita Bala, and Serge Belongie. 2015. Learning visual clothing style with heterogeneous dyadic co- occurrences. In Proceedings of the IEEE international conference on computer vision. 4642–4650

2015
[20]

Tianyu Wu, Shizhu He, Jingping Liu, Siqi Sun, Kang Liu, Qing-Long Han, and Yang Tang. 2023. A brief overview of ChatGPT: The history, status quo and potential future development. IEEE/CAA Journal of Automatica Sinica 10, 5 (2023), 1122–1136

2023
[21]

Ruobing Xie, Yanlei Liu, Shaoliang Zhang, Rui Wang, Feng Xia, and Leyu Lin
[22]

Siyuan Xing, Qiulei Dong, and Zhanyi Hu. 2022. SCE-Net: Self-and cross- enhancement network for single-view height estimation and semantic segmenta- tion. Remote Sensing 14, 9 (2022), 2252

2022
[23]

Fan Yang, Zheng Chen, Ziyan Jiang, Eunah Cho, Xiaojiang Huang, and Yanbin Lu. 2023. Palr: Personalization aware llms for recommendation. arXiv preprint arXiv:2305.07622 (2023)

2023 arXiv
[2021]

In Proceedings of the Web Conference 2021

Personalized approximate pareto-efficient recommendation. In Proceedings of the Web Conference 2021. 3839–3849

2021
[2023]

arXiv:2311.18702 [cs.CL]

CritiqueLLM: Scaling LLM-as-Critic for Effective and Explainable Evalua- tion of Large Language Model Generation. arXiv:2311.18702 [cs.CL]

arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Ziad Al-Halah, Rainer Stiefelhagen, and Kristen Grauman. 2017. Fashion for- ward: Forecasting visual style in fashion. In Proceedings of the IEEE international conference on computer vision . 388–397

work page 2017

[2] [2]

Victor Carbune, Hassan Mansoor, Fangyu Liu, Rahul Aralikatte, Gilles Baechler, Jindong Chen, and Abhanshu Sharma. 2024. Chart-based reasoning: Transferring capabilities from llms to vlms. arXiv preprint arXiv:2403.12596 (2024)

arXiv 2024

[3] [3]

Wen-Huang Cheng, Sijie Song, Chieh-Yun Chen, Shintami Chusnul Hidayati, and Jiaying Liu. 2021. Fashion Meets Computer Vision: A Survey. ACM Comput. Surv. 54, 4, Article 72 (jul 2021), 41 pages. https://doi.org/10.1145/3447239

work page doi:10.1145/3447239 2021

[4] [4]

Yujuan Ding, Zhihui Lai, PY Mok, and Tat-Seng Chua. 2023. Computational Technologies for Fashion Recommendation: A Survey. Comput. Surveys 56, 5 (2023), 1–45

2023

[5] [5]

Xintong Han, Zuxuan Wu, Yu-Gang Jiang, and Larry S Davis. 2017. Learning fashion compatibility with bidirectional lstms. In Proceedings of the 25th ACM international conference on Multimedia . 1078–1086

work page 2017

[6] [6]

Hyunwoo Hwangbo, Yang Sok Kim, and Kyung Jin Cha. 2018. Recommendation system development for fashion retail e-commerce.Electronic Commerce Research and Applications 28 (2018), 94–101

2018

[7] [7]

Pei Ke, Bosi Wen, Zhuoer Feng, Xiao Liu, Xuanyu Lei, Jiale Cheng, Shengyuan Wang, Aohan Zeng, Yuxiao Dong, Hongning Wang, Jie Tang, and Minlie Huang

work page

[8] [8]

Yen-Liang Lin, Son Tran, and Larry S. Davis. 2020. Fashion Outfit Complementary Item Retrieval. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE. https://doi.org/10.1109/cvpr42600.2020.00337

arXiv 2020