Pith. sign in

REVIEW 4 major objections 4 minor 81 references

Fine-tuning a vision encoder on synthetic-only data improves instance retrieval on six of seven benchmarks, without using any real training images.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 10:38 UTC pith:UOCWJUEX

load-bearing objection Genuinely new and mostly persuasive: synthetic-only instance-level training works, but the seed-as-instance assumption and a modified MET protocol keep it short of proven. the 4 major comments →

arxiv 2510.09171 v2 pith:UOCWJUEX submitted 2025-10-10 cs.CV

Instance-Level Generation for Representation Learning

Yankun Wu , Zakaria Laskar , Giorgos Kordopatis-Zilos , Noa Garcia , Giorgos Tolias This is my paper
classification cs.CV
keywords instance-level recognitionimage retrievalsynthetic training datadiffusion modelsmetric learningrepresentation learningfoundation modelsillumination relighting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

ILGen produces instance-level training data without a single real image: a large language model lists object categories for a target domain, a diffusion model generates one object per random seed, and an illumination model re-renders each object onto four new backgrounds. Each seed is treated as an instance, so the four relit versions are the instance's positive views. After fine-tuning a foundation vision encoder on this synthetic set with a retrieval-metric loss, average retrieval performance across seven benchmarks rises from 47.5 to 52.6 mAP, with gains on six of the seven datasets. The paper argues this makes domain names the only input needed to train an instance-level recognition model for that domain.

Core claim

The central claim is that a single representation model fine-tuned exclusively on synthetically generated instance-level data—constructed by treating each diffusion-model seed as an instance and relighting it onto four backgrounds—improves retrieval across seven instance-level recognition benchmarks, without using any real images for training. In the main comparison, the ILGen-ALL dataset (20K synthetic instances, half generic and half domain-specific) raises the average mAP of a pretrained SigLIP encoder from 47.5 to 52.6; the gains appear on all benchmarks except SOP, which is unchanged. The authors also show that the improvement is not tied to one backbone or loss: fine-tuning CLIP and Vi

What carries the argument

The load-bearing identity is the seed-as-instance rule: each random seed passed to the text-to-image model defines a distinct object instance, and relighting that single generated object onto N=4 synthesized backgrounds creates the instance's image set. This turns a batch of unlabeled synthetic images into a metric-learning dataset with automatically defined positives (same seed) and negatives (different seeds). The training objective is a recall@k surrogate loss, which optimizes the encoder to rank a query's same-seed views above all other images in large batches that simulate a retrieval task.

Load-bearing premise

Everything rests on the assumption that two different random seeds produce two different objects while relighting a single seed preserves the same object; if seeds duplicate or relighting changes identity, the synthetic positive/negative labels are wrong and training collapses into category-level discrimination.

What would settle it

Measure the margin between same-seed and different-seed similarities: feed a set of generated objects through the relighting step, then use a frozen instance-level descriptor to check whether the four relit versions of one seed are consistently closer to each other than to relit versions of other seeds. If the same-seed margin is not clearly positive—or if human raters cannot reliably group relit images by seed—the instance labels are unreliable and the retrieval gains should be attributed to category-level features rather than true instance discrimination.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A practitioner can produce a domain-specific ILR model from a domain name alone, skipping image collection, filtering, and manual instance annotations.
  • Synthetic instance data transfers to real test distributions well enough to lift retrieval on standard benchmarks, so it can complement or replace real data in ILR.
  • The benefit appears across different foundation backbones and different metric-learning losses, so the training recipe is not hostage to a single architecture.
  • Training on domain-specific synthetic objects boosts the corresponding benchmarks (artwork, landmark, product), while generic objects help multi-domain retrieval sets.
  • Even a small set of 5K synthetic instances improves over the pretrained backbone, and replacing diverse backgrounds with clean backgrounds lowers performance on cluttered test sets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the seed-as-instance rule holds, the same recipe could be applied to video or multi-camera settings, where one generated object relit and reposed across frames supplies natural temporal positives, extending the method beyond still-image retrieval.
  • The reported gains may partly come from the model learning to ignore background and illumination; a testable prediction is that per-query improvements concentrate on cases where the query and database differ mainly in background or lighting, as suggested by the clean-background ablation's drop on cluttered benchmarks.
  • A simple quality gate could make the pipeline stronger: automatically reject or re-label near-duplicate seeds (e.g., by perceptual similarity within a category) before training, which would reduce false negatives caused by two seeds depicting the same object.
  • Since the input is a text description, the method extends naturally to long-tail or fast-changing domains like new product lines or emerging landmarks, where curated training data lags behind the need.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes ILGen, a pipeline that generates synthetic instance-level training data without any real images. For a target domain (or a generic domain), an LLM produces object category names, Stable Diffusion Turbo generates images with different seeds (each seed treated as one object instance), ICLight relights and re-backgrounds each generated image N times to create positive views of the same instance, and a pretrained vision encoder (SigLIP, CLIP, ViT) is fine-tuned with the recall@k surrogate loss on these synthetic instance-level classes. The authors report that fine-tuning SigLIP on ILGen-ALL improves average mAP on seven instance-level retrieval benchmarks from 47.5 to 52.6, with gains on six of seven datasets, and that similar improvements hold for CLIP and ViT-B. They also compare against training on Objaverse renders and on real labeled images at matched sizes, report ablations over number of instances, backgrounds, LLMs, GDMs, losses, and background generators, and provide a qualitative leakage check.

Significance. If the claims hold, the paper makes a significant contribution: it offers a fully synthetic route to instance-level representation learning, replacing manual data collection with a pipeline whose only input is a domain description. The paper is commendably broad in evaluation: seven benchmarks, three backbones, several losses, multiple training-set variants, and comparison to real-image training at matched scale. The public release of code and models is a further strength. The ILGen-G variant, using only generic everyday-object categories, still improves most benchmarks, which supports transfer beyond the specific categories in the test domains. However, the central claim depends on the untested assumption that GDM seeds define distinct object instances and that ICLight relighting preserves instance identity; the paper's own S8 ablation shows that violating identity preservation is catastrophic. In addition, the default ILGen-ALL training set deliberately contains object categories from the evaluation domains, and the leakage check is only qualitative. These concerns are addressable with additional experiments, but they are load-bearing for the paper's central conclusion.

major comments (4)
  1. [§3.2 (Object instance generation) and Table 7] The entire instance-level supervisory signal is based on the assumption that 'generating images with different random seeds produces variations that are distinct and recognizable as separate instances within the same category,' and that the four ICLight relightings of a single seed are positives. No quantitative validation of either condition is provided. The importance of identity preservation is demonstrated by the paper's own ablation ID-S8: replacing ICLight with SD v2.0 for background generation drops average mAP to 47.1, below the pretrained model's 47.5, explicitly attributed to 'poor identity preservation.' This shows that the method's success is contingent on exactly the property that is only assumed. I request a quantitative identity-preservation evaluation (e.g., retrieval/verification accuracy on held-out synthetic pairs, or a human study on whether relightings depict the sam
  2. [§4.1 (Training set variants), Table 1, and §4.3 (Train and test set overlap)] ILGen-ALL, the default training set, contains 10K objects from the same three specific domains (art, landmark, product) used in six of the seven evaluation benchmarks. The domain-specific category lists are generated with prompts that mention exactly these domains, and examples such as 'temple' and 'French empire clock' appear close to test-domain content. The claimed 'no leakage' evidence in §4.3 is qualitative: the authors visually inspect the highest-similarity pairs (Figure 7). This does not rule out GDM memorization of test-domain images or category-level leakage. I request a quantitative leakage test, e.g., a membership/nearest-neighbor analysis over all training images with appropriate baselines, and/or an evaluation of ILGen-G alone on the three specific-domain benchmarks to show that gains do not require seeing categories from those domains. The ILGen-G results in Table 3 partia
  3. [§4.1 (Evaluation benchmarks, MET)] The MET benchmark is adapted in a nonstandard way: 'we retain only queries with at least one positive match in the database, i.e. we discard the distractor queries, and keep only the first positive per query.' This changes the task and can inflate mAP if hard queries or multiple positives are removed. The reported MET mAP@100 is therefore not directly comparable to published MET retrieval numbers. Please report results on the original MET retrieval protocol as well, or justify why the adapted protocol is preferable and show that the conclusion is unchanged under the standard protocol.
  4. [Tables 3, 5, 6, 7] All main results are single runs with no error bars or repeated-seed statistics. Several differences are small (e.g., SOP 55.4 vs. 55.3, GLDv2 15.7 vs. 18.3 in Table 3; several 0.1–0.3 differences in Table 7). Given the claim of 'consistent improvement', at least the key comparisons in Tables 3 and 7 should be repeated with multiple training seeds and reported with mean and standard deviation, or an appropriate significance test.
minor comments (4)
  1. [§4.2] The text says ILGen-ALL improves all datasets 'except SOP where performance does not change'; Table 3 shows SOP drops from 55.4 to 55.3. This is fine, but 'does not change' should be reworded to 'does not improve' or reported as a negligible drop.
  2. [Table 7] The row labels in Table 7 are somewhat terse; abbreviation 'pos' for number of images per instance is not defined in the table caption. Also, ID-S8 changes the background generation model to SD v2.0 while keeping the GDM as SD Turbo; clarifying that 'SD v2.0' in the background column is a different use of the same underlying model would help.
  3. [References] There are minor typographical issues, e.g., the author name 'Lipovsk` y' in reference [Ypsilantis et al. 2023] and 'Objaverse' spelled inconsistently in Tables S3/S10. These should be corrected in a final polish.
  4. [Figure 6] The caption says 'Points below the diagonal indicate a query with improved performance when fine-tuned on ILGen-ALL,' which is correct given the axis assignment, but the phrasing could be clarified to avoid ambiguity for the reader (below diagonal = ILGen-ALL mAP > pretrained mAP).

Circularity Check

0 steps flagged

No significant circularity: seed-based instance labels are explicit assumptions and benchmarks are external; only minor non-load-bearing self-citations.

full rationale

Walk-through: §3.2 defines training classes by GDM seed: 'We assume that generating images with different random seeds produces variations that are distinct and recognizable as separate instances within the same category. Therefore... each... is treated as a separate class.' This is an explicit modeling assumption used to construct the training set, not a derived prediction, and it is not what the benchmarks measure. Evaluation (§4.2, Table 3) is against external ILR datasets (MET, R-Oxford/R-Paris, GLDv2, SOP, INSTRE, mini-ILIAS) whose positives are real images of the same object, so the reported improvement (avg mAP 47.5 to 52.6) is an empirical result that could fail if the assumption were wrong, not a tautology. Including domain-specific categories in ILGen-ALL is a disclosed design choice and is controlled by ILGen-G (generic objects only), which still improves most benchmarks. Table 7 ID-S8 (ICLight to SD v2.0) is a validity probe showing identity preservation matters; it does not reduce the method to its own input. The leakage check (§4.3) is qualitative but is a post-hoc validity check, not a circular step. Self-citations to Patel et al. (recall@k loss), MET (Ypsilantis et al.), and ILIAS (Kordopatis-Zilos et al.) are references to public, externally available benchmarks and losses and are not used to prove the central claim. Thus no reduction-by-construction is present; the substantive concerns are about assumption validity, not circularity, so score 2 reflects only minor, non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on several domain assumptions about the reliability of generative models as label sources; no new physical entities or forces are introduced. The main free parameters are dataset-construction choices rather than fitted physical constants.

free parameters (5)
  • C object categories per domain = generic 2,000; art 200; landmark 50; product 200
    Hand-chosen counts of object categories per domain; controls dataset composition and matches instance budget, not derived from a principle.
  • K images per category = generic 10; art 15; landmark 80; product 15
    Hand-chosen number of seeds per category; larger K for landmarks gives more instances; affects class balance and total dataset size.
  • N backgrounds per instance = 4
    Number of ICLight backgrounds per generated object; ablations test 2, 3, and 4 but not a principled selection.
  • padding amount = random up to 50% of resolution
    Added before relighting to vary size and position; ablation S9 shows removing it drops average by 1.1.
  • ILGen-ALL mix ratio = 50% generic / 50% specific
    ILGen-ALL uses 10K generic + 10K domain-specific instances; chosen by hand after comparing domain-specific variants.
axioms (4)
  • domain assumption GDM seeds define distinct object instances; variations of one seed define the same instance.
    §3.2 Object instance generation. Label correctness of the whole synthetic dataset rests on this assumption; no quantitative validation is provided.
  • domain assumption ICLight background generation and relighting preserves the identity of the generated object.
    §3.2 Background generation. If identity drifts during relighting, the synthetic positives become false positives.
  • domain assumption LLM, GDM, and ICLight pretraining cover the visual content of the evaluation domains.
    §3.2/§4.1. The method's utility depends on generative models having seen relevant concepts; no independent evidence beyond benchmark gains.
  • domain assumption The recall@k surrogate loss is a valid optimization target for retrieval performance.
    §3.3 and Patel et al. 2022; training objective is only an approximation of the test-time retrieval metric.

pith-pipeline@v1.3.0-alltime-deepseek · 16692 in / 12166 out tokens · 536845 ms · 2026-08-04T10:38:15.068978+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Instance-Level Generation for Representation Learning." pith.science (2026). https://pith.science/paper/UOCWJUEX

@misc{pith2026251009171,
  author       = {Pith},
  title        = {Pith review of: Instance-Level Generation for Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UOCWJUEX}},
  note         = {Machine review of arXiv:2510.09171}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Instance-level recognition (ILR) focuses on identifying individual objects rather than broad categories, offering the highest granularity in image classification. However, this fine-grained nature makes creating large-scale annotated datasets challenging, limiting ILR's real-world applicability across domains. To overcome this, we introduce a novel approach that synthetically generates diverse object instances from multiple domains under varied conditions and backgrounds, forming a large-scale training set. Unlike prior work on automatic data synthesis, our method is the first to address ILR-specific challenges without relying on any real images. Fine-tuning foundation vision models on the generated data significantly improves retrieval performance across seven ILR benchmarks spanning multiple domains. Our approach offers a new, efficient, and effective alternative to extensive data collection and curation, introducing a new ILR paradigm where the only input is the names of the target domains, unlocking a wide range of real-world applications. The code and pretrained models are publicly available at https://github.com/yankungou/ILGen.

Figures

Figures reproduced from arXiv: 2510.09171 by Giorgos Kordopatis-Zilos, Giorgos Tolias, Noa Garcia, Yankun Wu, Zakaria Laskar.

Figure 1
Figure 1. Figure 1: Examples of images generated for learning instance-level representations. Given an object generated [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of instance-level training data generation. A domain name or description is the only [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of object instances generated by GDM for specific categories. We show the category [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Examples of object instances generated by GDM (column 1), and the generated images that leave [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training batch construction for instance-level representation learning. A batch simulates a retrieval [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Average Precision (AP) per query for the pretrained backbone (y-axis) and the backbone fine-tuned [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Pairs of ILGen-ALL and test sets with the highest similarity score. While these pairs share some [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

81 extracted references · 6 linked inside Pith

  1. [1]

    Augmented reality meets computer vision: Efficient data generation for urban driving scenes

    Hassan Abu Alhaija, Siva Karthik Mustikovela, Lars Mescheder, Andreas Geiger, and Carsten Rother. Augmented reality meets computer vision: Efficient data generation for urban driving scenes. IJCV, 2018

    2018

  2. [2]

    Unicom: Universal and compact representation learning for image retrieval

    Xiang An, Jiankang Deng, Kaicheng Yang, Jaiwei Li, Ziyong Feng, Jia Guo, Jing Yang, and Tongliang Liu. Unicom: Universal and compact representation learning for image retrieval. In ICLR, 2023

    2023

  3. [3]

    This dataset does not exist: training models from generated images

    Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick P \'e rez. This dataset does not exist: training models from generated images. In ICASSP, 2020

    2020

  4. [4]

    Large scale gan training for high fidelity natural image synthesis

    Andrew Brock. Large scale gan training for high fidelity natural image synthesis. In ICLR, 2018

    2018

  5. [5]

    Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation

    Dingding Cai, Janne Heikkil \"a , and Esa Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In CVPR, 2022

    2022

  6. [6]

    Unifying deep local and global features for image search

    Bingyi Cao, Andr \'e Araujo, and Jack Sim. Unifying deep local and global features for image search. In ECCV, 2020

    2020

  7. [7]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015

    Pith/arXiv arXiv 2015

  8. [8]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020

    2020

  9. [9]

    Pali: A jointly-scaled multilingual language-image model

    Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. In ICLR, 2023

    2023

  10. [10]

    Learning semantic segmentation from synthetic data: A geometrically guided input-output adaptation approach

    Yuhua Chen, Wen Li, Xiaoran Chen, and Luc Van Gool. Learning semantic segmentation from synthetic data: A geometrically guided input-output adaptation approach. In CVPR, 2019

    2019

  11. [11]

    Learning a similarity metric discriminatively, with application to face verification

    Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In CVPR, 2005

    2005

  12. [12]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In CVPR, 2023

    2023

  13. [13]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019

    2019

  14. [14]

    ImageNet : A large-scale hierarchical image database

    Wei Dong, Richard Socher, Li Li-Jia, Kai Li, and Li Fei-Fei. ImageNet : A large-scale hierarchical image database. In CVPR, 2009

    2009

  15. [15]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021

    2021

  16. [16]

    Cut, paste and learn: Surprisingly easy synthesis for instance detection

    Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In ICCV, 2017

    2017

  17. [17]

    The group loss for deep metric learning

    Ismail Elezi, Sebastiano Vascon, Alessandro Torcinovich, Marcello Pelillo, and Laura Leal-Taix \'e . The group loss for deep metric learning. In ECCV, 2020

    2020

  18. [18]

    Scaling laws of synthetic images for model training

    Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024

    2024

  19. [19]

    Instructdiffusion: A generalist modeling interface for vision tasks

    Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In CVPR, 2024

    2024

  20. [20]

    Synthesizing training data for object detection in indoor scenes

    Georgios Georgakis, Arsalan Mousavian, Alexander C Berg, and Jana Kosecka. Synthesizing training data for object detection in indoor scenes. arXiv preprint arXiv:1702.07836, 2017

    Pith/arXiv arXiv 2017

  21. [21]

    Revisiting the fisher vector for fine-grained classification

    Philippe-Henri Gosselin, Naila Murray, Herv\'e J\'egou, and Florent Perronnin. Revisiting the fisher vector for fine-grained classification. Pattern Recognition Letters, 2014

    2014

  22. [22]

    Dimensionality reduction by learning an invariant mapping

    Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR, 2006

    2006

  23. [23]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016

    2016

  24. [24]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020

    2020

  25. [25]

    Local descriptors optimized for average precision

    Kun He, Yan Lu, and Stan Sclaroff. Local descriptors optimized for average precision. In CVPR, 2018

    2018

  26. [26]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Proxy anchor loss for deep metric learning

    Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Proxy anchor loss for deep metric learning. In CVPR, 2020

    2020

  28. [28]

    Self-taught metric learning without labels

    Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Self-taught metric learning without labels. In CVPR, 2022

    2022

  29. [29]

    Adam: A method for stochastic optimization

    Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015

    2015

  30. [30]

    ILIAS : Instance-level image retrieval at scale

    Giorgos Kordopatis-Zilos, Vladan Stojni \'c , Anna Manko, Pavel S uma, Nikolaos-Antonios Ypsilantis, Nikos Efthymiadis, Zakaria Laskar, Ji r \' Matas, Ond r ej Chum, and Giorgos Tolias. ILIAS : Instance-level image retrieval at scale. In CVPR, 2025

    2025

  31. [31]

    Cross-image-attention for conditional embeddings in deep metric learning

    Dmytro Kotovenko, Pingchuan Ma, Timo Milbich, and Bj \"o rn Ommer. Cross-image-attention for conditional embeddings in deep metric learning. In CVPR, 2023

    2023

  32. [32]

    Fine-grained recognition without part annotations

    Jonathan Krause, Hailin Jin, Jianchao Yang, and Li Fei-Fei. Fine-grained recognition without part annotations. In CVPR, 2015

    2015

  33. [33]

    Cosypose: Consistent multi-view multi-object 6d pose estimation

    Yann Labb \'e , Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypose: Consistent multi-view multi-object 6d pose estimation. In ECCV, 2020

    2020

  34. [34]

    Correlation verification for image retrieval

    Seongwon Lee, Hongje Seong, Suhyeon Lee, and Euntai Kim. Correlation verification for image retrieval. In CVPR, 2022

    2022

  35. [35]

    Syncdreamer: Generating multiview-consistent images from a single-view image

    Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024

    2024

  36. [36]

    Deepfashion: Powering robust clothes recognition and retrieval with rich annotations

    Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaoou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016

    2016

  37. [37]

    Amass: Archive of motion capture as surface shapes

    Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019

    2019

  38. [38]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023

    2023

  39. [39]

    A metric learning reality check

    Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In ECCV, 2020

    2020

  40. [40]

    Deep metric learning via lifted structured feature embedding

    Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016

    2016

  41. [41]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth \'e e Darcet, Th \'e o Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  42. [42]

    Recall@ k surrogate loss with large batches and similarity mixup

    Yash Patel, Giorgos Tolias, and Ji r \' Matas. Recall@ k surrogate loss with large batches and similarity mixup. In CVPR, 2022

    2022

  43. [43]

    RP2K : A large-scale retail product dataset for fine-grained image classification

    Jingtian Peng, Chang Xiao, and Yifan Li. RP2K : A large-scale retail product dataset for fine-grained image classification. arXiv preprint arXiv:2006.12634, 2020

    Pith/arXiv arXiv 2006

  44. [44]

    Learning deep object detectors from 3d models

    Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object detectors from 3d models. In ICCV, 2015

    2015

  45. [45]

    Philbin, O

    J. Philbin, O. Chum, M. Isard, J. Sivic, and A. Zisserman. Object retrieval with large vocabularies and fast spatial matching. In CVPR, 2007

    2007

  46. [46]

    Philbin, O

    J. Philbin, O. Chum, M. Isard, J. Sivic, and A. Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In CVPR, 2008

    2008

  47. [47]

    Softtriple loss: Deep metric learning without triplet sampling

    Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In ICCV, 2019

    2019

  48. [48]

    Revisiting oxford and paris: Large-scale image retrieval benchmarking

    Filip Radenovi \'c , Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ond r ej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018

    2018

  49. [49]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021

    2021

  50. [50]

    Dreambooth3d: Subject-driven text-to-3d generation

    Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In ICCV, 2023

    2023

  51. [51]

    Robust and decomposable average precision for image retrieval

    Elias Ramzi, Nicolas Thome, Cl \'e ment Rambour, Nicolas Audebert, and Xavier Bitot. Robust and decomposable average precision for image retrieval. In NeurIPS, 2021

    2021

  52. [52]

    Hierarchical average precision training for pertinent image retrieval

    Elias Ramzi, Nicolas Audebert, Nicolas Thome, Cl \'e ment Rambour, and Xavier Bitot. Hierarchical average precision training for pertinent image retrieval. In ECCV, 2022

    2022

  53. [53]

    Learning with average precision: Training image retrieval with a listwise loss

    Jerome Revaud, Jon Almaz \'a n, Rafael S Rezende, and Cesar Roberto de Souza. Learning with average precision: Training image retrieval with a listwise loss. In ICCV, 2019

    2019

  54. [54]

    Optimizing rank-based metrics with blackbox differentiation

    Michal Rol \' nek, V \' t Musil, Anselm Paulus, Marin Vlastelica, Claudio Michaelis, and Georg Martius. Optimizing rank-based metrics with blackbox differentiation. In CVPR, 2020

    2020

  55. [55]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022

    2022

  56. [56]

    The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes

    German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016

    2016

  57. [57]

    Revisiting training strategies and generalization performance in deep metric learning

    Karsten Roth, Timo Milbich, Samarth Sinha, Prateek Gupta, Bjorn Ommer, and Joseph Paul Cohen. Revisiting training strategies and generalization performance in deep metric learning. In ICML, 2020

    2020

  58. [58]

    On rendering synthetic images for training an object detector

    Artem Rozantsev, Vincent Lepetit, and Pascal Fua. On rendering synthetic images for training an object detector. CVIU, 2015

    2015

  59. [59]

    Imagenet large scale visual recognition challenge

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015

    2015

  60. [60]

    Fake it till you make it: Learning transferable representations from synthetic imagenet clones

    Mert B \"u lent Sar y ld z, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In CVPR, 2023

    2023

  61. [61]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024

    2024

  62. [62]

    Facenet: A unified embedding for face recognition and clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, 2015

    2015

  63. [63]

    LAION-400M : Open dataset of clip-filtered 400 million image-text pairs

    Christoph Schuhmann, Robert Kaczmarczyk, Aran Komatsuzaki, Aarush Katta, Richard Vencu, Romain Beaumont, Jenia Jitsev, Theo Coombes, and Clayton Mullis. LAION-400M : Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshop Datacentric AI, 2021

    2021

  64. [64]

    Learning intra-batch connections for deep metric learning

    Jenny Seidenschwarz, Ismail Elezi, and Laura Leal-Taix \'e . Learning intra-batch connections for deep metric learning. In ICML, 2021

    2021

  65. [65]

    Global features are all you need for image retrieval and reranking

    Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, Andr\'e Araujo, and Bingyi Cao. Global features are all you need for image retrieval and reranking. In ICCV, 2023

    2023

  66. [66]

    Improved deep metric learning with multi-class n-pair loss objective

    Kihyuk Sohn. Improved deep metric learning with multi-class n-pair loss objective. In NeurIPS, 2016

    2016

  67. [67]

    Ames: Asymmetric and memory-efficient similarity estimation for instance-level retrieval

    Pavel Suma, Giorgos Kordopatis-Zilos, Ahmet Iscen, and Giorgos Tolias. Ames: Asymmetric and memory-efficient similarity estimation for instance-level retrieval. In ECCV, 2024

    2024

  68. [68]

    Personalized representation from personalized generation

    Shobhita Sundaram, Julia Chae, Yonglong Tian, Sara Beery, and Phillip Isola. Personalized representation from personalized generation. In ICLR, 2025

    2025

  69. [69]

    Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis

    Eu Wern Teh, Terrance DeVries, and Graham W Taylor. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In ECCV, 2020

    2020

  70. [70]

    Stablerep: Synthetic images from text-to-image models make strong visual representation learners

    Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. In NeurIPS, 2024

    2024

  71. [71]

    Self6d: Self-supervised monocular 6d object pose estimation

    Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In ECCV, 2020

    2020

  72. [72]

    Cosface: Large margin cosine loss for deep face recognition

    Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In CVPR, 2018

    2018

  73. [73]

    Instre: a new benchmark for instance-level object retrieval and recognition

    Shuang Wang and Shuqiang Jiang. Instre: a new benchmark for instance-level object retrieval and recognition. ACM TOMM, 2015

    2015

  74. [74]

    Google landmarks dataset v2 - A large-scale benchmark for instance-level recognition and retrieval

    Tobias Weyand, Andr \' e Araujo, Bingyi Cao, and Jack Sim. Google landmarks dataset v2 - A large-scale benchmark for instance-level recognition and retrieval. In CVPR, 2020

    2020

  75. [75]

    Not only generative art: Stable diffusion for content-style disentanglement in art analysis

    Yankun Wu, Yuta Nakashima, and Noa Garcia. Not only generative art: Stable diffusion for content-style disentanglement in art analysis. In ICMR, 2023

    2023

  76. [76]

    The met dataset: Instance-level recognition for artworks

    Nikolaos-Antonios Ypsilantis, Noa Garcia, Guangxing Han, Sarah Ibrahimi, Nanne Van Noord, and Giorgos Tolias. The met dataset: Instance-level recognition for artworks. In NeurIPS, 2021

    2021

  77. [77]

    Towards universal image embeddings: A large-scale dataset and challenge for generic image representations

    Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Bingyi Cao, M \'a rio Lipovsk \`y , Pelin Dogan-Sch \"o nberger, Grzegorz Makosa, Boris Bluntschli, Mojtaba Seyedhosseini, Ond r ej Chum, and Andr \'e Araujo. Towards universal image embeddings: A large-scale dataset and challenge for generic image representations. In ICCV, 2023

    2023

  78. [78]

    Classification is a strong baseline for deep metric learning

    Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. arXiv preprint arXiv:1811.12649, 2018

    Pith/arXiv arXiv 2018

  79. [79]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023

    2023

  80. [80]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023

    2023

Showing first 80 references.