3rd Place Solution to "Google Landmark Retrieval 2020"

Mei, Ke; Li, Lei; Xu, Jinchang; Cheng, Yanhua; Lin, Yugeng

doi:10.48550/arxiv.2008.10480

Cited by 4 publications

(4 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Given an embedding model ϕ : x → R D , the task of image retrieval is to identify the object/content of interest in gallery resorting to the query-gallery similarities, i.e., ∥ϕ(Q) − ϕ(G)∥. Following state-of-the-art metric learning methods (Jeon 2020;Mei et al 2020) in image retrieval, we employ classification as a pretext task in the form of ArcFace loss (Deng et al 2019).…”

Section: Image Retrievalmentioning

confidence: 99%

“…We demonstrate the superiority of DMU against the other model upgrading methods on multiple widelyacknowledged image retrieval benchmarks, including the large-scale landmark retrieval and face recognition. Following state-of-the-art methods (Jeon 2020;Mei et al 2020), we perform model upgrades on GLDv2-train (Weyand et al 2020) and MS1Mv3 (Deng et al 2019), and evaluate retrieval on GLDv2-test, ROxford (Radenović et al 2018b), RParis, and IJB-C (Radenović et al 2018a) respectively. DMU consistently surpasses the competitors in terms of both new-to-new and new-to-old retrieval performances.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Darwinian Model Upgrades: Model Evolving with Selective Compatibility

Zhang

et al. 2023

AAAI

View full text Add to dashboard Cite

The traditional model upgrading paradigm for retrieval requires recomputing all gallery embeddings before deploying the new model (dubbed as "backfilling"), which is quite expensive and time-consuming considering billions of instances in industrial applications. BCT presents the first step towards backward-compatible model upgrades to get rid of backfilling. It is workable but leaves the new model in a dilemma between new feature discriminativeness and new-to-old compatibility due to the undifferentiated compatibility constraints. In this work, we propose Darwinian Model Upgrades (DMU), which disentangle the inheritance and variation in the model evolving with selective backward compatibility and forward adaptation, respectively. The old-to-new heritable knowledge is measured by old feature discriminativeness, and the gallery features, especially those of poor quality, are evolved in a lightweight manner to become more adaptive in the new latent space. We demonstrate the superiority of DMU through comprehensive experiments on large-scale landmark retrieval and face recognition benchmarks. DMU effectively alleviates the new-to-new degradation at the same time improving new-to-old compatibility, rendering a more proper model upgrading paradigm in large-scale retrieval systems.Code: https://github.com/TencentARC/OpenCompatible.

show abstract

Section: Image Retrievalmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Darwinian Model Upgrades: Model Evolving with Selective Compatibility

Zhang

et al. 2023

AAAI

View full text Add to dashboard Cite

show abstract

“…λ is the loss weight coefficient for backward compatibility. Specifically, following state-of-the-art metric learning methods [14,22] in image retrieval, we adopt ArcFace loss [6] as the basic loss (L base = arc (φ new , ω new )), that is,…”

Section: Bidirectional Compatible Training (Bict)mentioning

confidence: 99%

Privacy-Preserving Model Upgrades with Bidirectional Compatible Training in Image Retrieval

Su¹,

Zhang²,

Ge³

et al. 2022

Preprint

View full text Add to dashboard Cite

The task of privacy-preserving model upgrades in image retrieval desires to reap the benefits of rapidly evolving new models without accessing the raw gallery images. A pioneering work [29] introduced backward-compatible training, where the new model can be directly deployed in a backfill-free manner, i.e., the new query can be directly compared to the old gallery features. Despite a possible solution, its improvement in sequential model upgrades is gradually limited by the fixed and under-quality old gallery embeddings. To this end, we propose a new model upgrade paradigm, termed Bidirectional Compatible Training (BiCT), which will upgrade the old gallery embeddings by forward-compatible training towards the embedding space of the backward-compatible new model. We conduct comprehensive experiments to verify the prominent improvement by BiCT and interestingly observe that the inconspicuous loss weight of backward compatibility actually plays an essential role for both backward and forward retrieval performance. To summarize, we introduce a new and valuable problem named privacy-preserving model upgrades, with a proper solution BiCT. Several intriguing insights are further proposed to get the most out of our method. * Equal contribution. Work done during Binjie's internship at ARC Lab, Tencent PCG. † Corresponding author. ‡ The conventional model upgrades in a retrieval system require to re-extract the gallery embeddings from raw images with the new model (dubbed as "backfill") before deployment.

show abstract

“…Interestingly, the advances in state-of-the-art image classification models over the past year, especially Efficient-Net [12], together with their open source ImageNet pretrained weights, have pushed the performance of global feature only models on this task so high that the additional benefit of local feature models could be marginal or negligible. This is evidenced by the fact that none of top three solutions in Retrieval 2020 [8, 14,10] nor the first place solution in Recognition 2020 [7] used local features.…”

Section: Introductionmentioning

confidence: 99%

Google Landmark Recognition 2020 Competition Third Place Solution

Ha,

Liu,

Liu

et al. 2020

Preprint

View full text Add to dashboard Cite

We present our third place solution to the Google Landmark Recognition 2020 competition. It is an ensemble of global features only Sub-center ArcFace models. We introduce dynamic margins for ArcFace loss, a family of tune-able margin functions of class size, designed to deal with the extreme imbalance in GLDv2 dataset. Progressive finetuning and careful postprocessing are also key to the solution. Our two submissions scored 0.6344 and 0.6289 on private leaderboard, both ranking third place out of 736 teams.

show abstract

3rd Place Solution to "Google Landmark Retrieval 2020"

Cited by 4 publications

References 20 publications

Darwinian Model Upgrades: Model Evolving with Selective Compatibility

Darwinian Model Upgrades: Model Evolving with Selective Compatibility

Privacy-Preserving Model Upgrades with Bidirectional Compatible Training in Image Retrieval

Google Landmark Recognition 2020 Competition Third Place Solution

Contact Info

Product

Resources

About