Pith. sign in

REVIEW 4 major objections 4 minor 37 references

Bridging Annotation Gaps: Transferring Labels to Align Object Detection Datasets

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proposes Label-Aligned Transfer (LAT), a data-centric framework that projects the annotations of multiple object detection datasets into a fixed target label space, claiming consistent target-domain gains of up to +4.8 AP…

desk verdict Plausible idea with consistent gains, but the missing loss and lack of direct label-transfer evaluation leave the central mechanism unverified. read the letter →

arxiv 2506.04737 v2 pith:LH7HEY3T submitted 2025-06-05 cs.CV

Mikhail Kennerley , Angelica Aviles-Rivero , Carola-Bibiane Schönlieb , Robby T. Tan This is my paper
classification cs.CV
keywords objectdetectionlabeltransferdatasetalignmentpseudo-labelingmulti-datasetlearningsemanticfeaturefusionspacedomainadaptation
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 argues that heterogeneous object detection datasets can be combined without merging their label ontologies: instead, source annotations are projected into a target dataset's label space. LAT trains one detector per dataset, generates pseudo-labels from each detector on every other dataset, and feeds these together with ground-truth boxes through a privileged proposal generator and a confidence-weighted attention module. The claimed payoff is consistent target-domain improvement, with +4.2 AP on datasets of mismatched class granularity and +4.8 AP when small datasets are augmented with large ones, all while preserving the target label space. A practitioner would care because this offers a path to leverage external detection data without re-annotating or changing the label scheme of a small task-specific dataset.

What carries the argument

The load-bearing mechanism is the pair of modules PPG and SFF. PPG substitutes for the region proposal network, using lightly augmented ground-truth boxes and pseudo-labels from the other datasets' detectors as region proposals. SFF computes scaled dot-product attention over RoI features, multiplies the similarity matrix by a confidence vector (1 for ground-truth proposals, the maximum classification score for pseudo-label proposals), and applies a row-wise clamp at $1/\sqrt{N}$ before fusing class-aware and region-feature value matrices. A frozen vision foundation model supplies the shared image features, and classification logits are masked to the target label space at inference.

What would settle it

Run LAT on a source-target pair where the source detector's pseudo-labels are known to be systematically wrong (for example a class the source detector never saw, or a heavy visual domain shift) and compare downstream AP to a target-only baseline; separately permute the pseudo-label class assignments and check whether AP collapses.

Watch

Extended reading notes

Core claim

LAT's central claim is that the $N(N-1)$ cross-dataset pseudo-label mappings produced by dataset-specialist detectors provide sufficient implicit semantic and spatial alignment to supervise a many-to-one label transfer into a chosen target label space. The framework replaces the region proposal network with a Privileged Proposal Generator that supplies ground-truth and pseudo-label boxes to RoI pooling, and a Semantic Feature Fusion module that attends over overlapping proposals, weighting ground-truth proposals with confidence 1 and pseudo-label proposals by their classification score. This lets the model learn relations between differently named or differently granular classes and different bounding-box conventions without paired annotations or a unified taxonomy. Experiments report downstream AP gains over a target-only baseline and over semi-supervised baselines across the tested benchmarks.

Load-bearing premise

Everything rests on the pseudo-labels produced by a detector trained on one dataset being accurate enough, when applied to another dataset's images, to serve as trustworthy cross-space supervision.

Editorial extensions

If this is right

  • Combining datasets with different label granularity becomes possible without hand-built taxonomies: the paper shows target-domain gains on Cityscapes, nuImages, and Waymo.
  • Small datasets can be augmented with large external datasets: ACDC and Cityscapes gain in the small-large benchmark, and longer training lets the larger datasets recover their performance.
  • Ground-truth-anchored transfer avoids the error propagation that the paper observes in student-teacher baselines under domain gap.
  • A fixed target ontology is preserved at inference because logits are masked to the target label space.
  • Even in a simple one-to-one transfer protocol, LAT matches the specialized LGPL method's reported AP on Synscapes to Cityscapes.

Reading between the lines

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

  • An extension the authors do not test: once trained on N datasets, LAT could serve as a reusable label translator that projects an unseen dataset into the target space, since its attention is over proposals rather than tied to the original dataset identities.
  • The method's quadratic $N(N-1)$ pseudo-label mappings suggest a scaling bottleneck; a natural extension would be to subsample source detectors or distill the mappings, but the paper does not analyze training cost.
  • The confidence-weighted SFF attention predicts that miscalibrated pseudo-label scores should hurt transfer; calibrating the source detectors' confidences before fusion could yield further gains beyond the reported results.
  • The gains concentrate on low-data target datasets, so LAT could plausibly combine with active learning to choose which external images to transfer first, though this is not explored in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes Label-Aligned Transfer (LAT), a data-centric framework that aims to transfer annotations from multiple source object detection datasets into the label space of a designated target dataset. LAT first trains one detector per dataset, uses these detectors to generate cross-space pseudo-labels, and then introduces two components into a two-stage detector: a Privileged Proposal Generator (PPG), which injects ground-truth and pseudo-label proposals in place of the RPN, and a Semantic Feature Fusion (SFF) module, which refines region features using class-aware attention over overlapping proposals. The method is evaluated on two benchmarks: Cityscapes/nuImages/Waymo for class-granularity mismatch and Cityscapes/ACDC/BDD100K/SHIFT for small-to-large dataset transfer, with additional experiments on Synscapes-to-Cityscapes. The authors report consistent target-domain AP improvements over semi-supervised baselines, with claimed gains of +4.2 AP and +4.8 AP.

Significance. LAT addresses a practically important problem: combining heterogeneous detection datasets while preserving a fixed target label space, without manual taxonomy merging. If the mechanism works as described, the framework would be a useful data-centric alternative to model-centric multi-dataset detectors. The paper's strengths include consistent AP gains across two benchmarks, ablations of the clamping and training strategies, and a comparison to LGPL on a simpler transfer protocol. However, the central claim of semantic and spatial label transfer is currently under-evidenced: the training objective of the LAT model is never specified, no direct evaluation of the transferred labels is reported, and the downstream AP numbers cannot distinguish genuine label transfer from self-training or from the extra ground-truth proposals injected by PPG. No code is released and all tables report single runs without error bars, which limits reproducibility.

major comments (4)
  1. [Section 3.2 and Section 4.2] The training objective of the LAT refinement model is never specified. No loss function, label assignment rule, or gradient equation is provided; the only hint is the Figure 2 caption, which says the classification output is filtered to compute loss on only the current dataset's label space. On a source image, the current label space is the source label space, not the target, so it is unclear how target-space supervision enters. If the loss on source images is to match the pseudo-labels of the target detector M_target, then SFF only re-weights those pseudo-labels and cannot recover classes that M_target missed; if the loss is to match source ground truths, then the model is not trained to transfer to the target at all. Additionally, Section 4.2 says the downstream detector is trained with a weighted cross-entropy loss using pseudo-label confidence weights, but it is not stated whether the training labels are LAT's refined outputs or the raw pseudo-labels. This gap directly undermines the central label-transfer claim.
  2. [Section 4.3, Tables 1 and 2] No direct evaluation of the transferred labels is reported. The only evidence for label transfer is downstream target-domain AP. A detector trained with source ground-truth boxes injected by PPG, or trained on a larger aggregated set, could achieve similar AP gains without learning any semantic or spatial correspondence. The authors should report precision, recall, or mAP of the LAT output labels against target-space annotations on a held-out subset, or at least include an ablation that removes cross-space pseudo-labels while retaining PPG, to isolate the transfer mechanism.
  3. [Section 3.2.2 and Table 4] The SFF attention mechanism is underspecified. The text says each row of the similarity matrix A is scaled such that its maximum value equals T = 1/sqrt(N), but then says clamping ensures each row's maximum does not exceed T; these are different operations, and the displayed equations do not resolve the ambiguity because the clamp bounds are not defined. Moreover, T is a hand-chosen value whose ablation only compares 1/N with 1/sqrt(N), so the claim that stronger scaling causes a substantial drop is not demonstrated by the reported table.
  4. [Section 4.2 and Section 1] The main benchmarks compare LAT only to Baseline, Student-Teacher, and Pseudo-Label. No comparison is made to representative multi-dataset detection methods such as ScaleDet, UniDet, or Plain-DET, and LGPL is compared only on the simpler Synscapes-to-Cityscapes protocol using numbers taken from the original paper. The phrase "state-of-the-art baselines" in Section 1 is therefore not supported. In addition, all tables report a single run without error bars or repeated seeds, and the code is not released, which limits the reader's ability to assess the reliability of the reported gains.
minor comments (4)
  1. [Abstract and Section 4.3] The abstract and introduction claim a "+4.2 AP" gain for the high-low class alignment setting, but Table 1 shows the largest gain in that setting is +4.4 (Cityscapes) and the mean gain is approximately +3.3; please correct the reported number.
  2. [Section 4.3] The statement that student-teacher approaches "consistently underperform, even falling below the baseline" is contradicted by Table 1 (nuImages: 40.1 vs. 39.2) and Table 2 (ACDC: 48.2 vs. 45.0); please revise this claim.
  3. [Section 3.1.2 and Section 4.2] The pseudo-label score threshold and the confidence thresholding values are never specified, although Section 3.1.2 says pseudo-labels are filtered by non-maximum suppression and score thresholding; these hyperparameters should be reported for reproducibility.
  4. [Throughout] There are several typographical issues, including "Tsinghua Univeristy" in the author affiliation, "generaliable" in Section 4.3, and "current datasets label space" in the Figure 2 caption. In addition, the key equations in Section 3.2.2 are unnumbered and introduced only in prose; numbering them would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LAT is an empirical pipeline evaluated on held-out target-domain AP; no prediction reduces to a fitted parameter or a load-bearing self-citation.

full rationale

The paper's derivation chain is empirical rather than mathematical. The claimed contribution, transferring source annotations into a target label space, is implemented by training dataset-specific detectors, generating pseudo-labels, and training a two-stage detector with PPG proposals and SFF attention. The downstream AP numbers in Tables 1, 2, 3, 5, and 6 are measured on target evaluation sets, not on the LAT training objective, so no fitted value is renamed as a prediction. The clamping threshold T = 1/sqrt(N) is an ablated hyperparameter, not derived from the target result. The authors' prior work (2PCNet, CAT) appears only as related work and semi-supervised baselines and is not load-bearing for the label-transfer claim. The closest concern is that for source images there is no ground-truth annotation in the target label space, so the only target-space signal is pseudo-labels from the target detector; however, the paper explicitly acknowledges the absence of paired supervision in Section 3.1.1, and the reported gains are empirical. This is an evidence/completeness limitation, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the reliability of cross-dataset pseudo-labels, the transferability of frozen DINOv2 features, and the assumption that overlapping proposals provide enough correspondence signal. These are domain assumptions about the data and the learned representations, not circular definitions. The only hand-tuned free parameters are the attention threshold and the PPG augmentation rates, plus an unspecified pseudo-label confidence threshold.

free parameters (3)
  • SFF attention scaling threshold T = 1/sqrt(N), where N is the number of datasets
    Section 3.2.2 and Table 4: the paper ablates clamping and scaling at 1/N versus 1/sqrt(N) and selects 1/sqrt(N) as the best value on the target benchmarks. This constant directly controls attention sharpness and changes AP by several points.
  • PPG ground-truth augmentation rates = jittering rate 0.5, removal rate 0.05
    Section 4.2 states these rates without sensitivity analysis. They control how much noisy privileged information is injected into the proposal generator and could materially affect the learned correspondences.
  • Pseudo-label confidence threshold = not specified in the paper
    Section 3.1.2 says pseudo-labels are filtered by score thresholding, but the threshold value is never reported. The quality of all downstream cross-space supervision depends on this choice.
assumptions (4)
  • domain assumption A single detector per dataset, trained on its native label space, produces reliable pseudo-labels in every other label space.
    Section 3.1.2: 'For each dataset D_n, we train a corresponding object detector M_n... generate pseudo-labels for each dataset under every other dataset's label space.' The entire transfer pipeline relies on these pseudo-labels being accurate enough to supervise semantic and spatial alignment.
  • domain assumption Frozen DINOv2 features provide a shared visual representation in which semantically similar objects across datasets are close enough for attention to discover correspondences.
    Section 3.2: 'we replace the conventional feature extractor with a frozen Vision Foundation Model (VFM), such as DINOv2.' The method does not fine-tune the backbone, so the success of cross-dataset feature fusion depends on the pretrained features transferring across domains.
  • domain assumption Overlapping proposals across label spaces correspond to the same physical objects, so attention over overlapping proposals can align classes.
    Section 3.2.1: 'these labels are derived from multiple label spaces, they often contain overlapping objects across datasets.' If pseudo-label boxes and ground-truth boxes do not overlap for the same object, the SFF attention mechanism has no signal to connect the differing class names.
  • domain assumption Ground-truth and pseudo-label proposals are sufficient to replace the region proposal network for the target label space.
    Section 3.2: PPG 'replaces the region proposal network' and supplies proposals to the RoI layer. If an object is missed by both the source detectors and the target ground truth, it will never be proposed, so the downstream detector cannot learn to detect it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Annotation Gaps: Transferring Labels to Align Object Detection Datasets." pith.science (2026). https://pith.science/paper/LH7HEY3T

@misc{pith2026250604737,
  author       = {Pith},
  title        = {Pith review of: Bridging Annotation Gaps: Transferring Labels to Align Object Detection Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LH7HEY3T}},
  note         = {Machine review of arXiv:2506.04737}
}
read the original abstract

Combining multiple object detection datasets offers a path to improved generalisation but is hindered by inconsistencies in class semantics and bounding box annotations. Some methods to address this assume shared label taxonomies and address only spatial inconsistencies; others require manual relabelling, or produce a unified label space, which may be unsuitable when a fixed target label space is required. We propose Label-Aligned Transfer (LAT), a label transfer framework that systematically projects annotations from diverse source datasets into the label space of a target dataset. LAT begins by training dataset-specific detectors to generate pseudo-labels, which are then combined with ground-truth annotations via a Privileged Proposal Generator (PPG) that replaces the region proposal network in two-stage detectors. To further refine region features, a Semantic Feature Fusion (SFF) module injects class-aware context and features from overlapping proposals using a confidence-weighted attention mechanism. This pipeline preserves dataset-specific annotation granularity while enabling many-to-one label space transfer across heterogeneous datasets, resulting in a semantically and spatially aligned representation suitable for training a downstream detector. LAT thus jointly addresses both class-level misalignments and bounding box inconsistencies without relying on shared label spaces or manual annotations. Across multiple benchmarks, LAT demonstrates consistent improvements in target-domain detection performance, achieving gains of up to +4.8AP over semi-supervised baselines.

Figures

Figures reproduced from arXiv: 2506.04737 by the authors.

Figure 1
Figure 1. Annotation discrepancies between three road-based object detection datasets: Cityscapes, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the LAT architecture. Dataset-specific pseudo-labels and ground-truth annota [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Comparison of labels on a Waymo im￾age. Ground-Truth labels are the original Waymo labels. Pseudo-labels and Label-Aligned Transfer are in the Cityscapes label space. Label-Aligned Transfer accurately labels the cyclist and bicycle into their respective classes and boxes. Privileged ground-truth information also results in more ac￾curate detections of smaller objects missing in the pseudo-labels. To enhance the fusi… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: We visualise the labelling hierarchy of vehicle classes in the Waymo, Cityscapes, and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [1]

    Scaledet: A scalable multi-dataset object detector

    Yanbei Chen, Manchen Wang, Abhay Mittal, Zhenlin Xu, Paolo Favaro, Joseph Tighe, and Davide Modolo. Scaledet: A scalable multi-dataset object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7288–7297, June 2023

    work page 2023

  2. [2]

    Towards universal object detection by domain attention

    Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards universal object detection by domain attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    work page 2019

  3. [3]

    Transferring labels to solve annotation mismatches across object detection datasets

    Yuan-Hong Liao, David Acuna, Rafid Mahmood, James Lucas, Viraj Uday Prabhu, and Sanja Fidler. Transferring labels to solve annotation mismatches across object detection datasets. In The Twelfth International Conference on Learning Representations , 2024. URL https: //openreview.net/forum?id=ChHx5ORqF0

    work page 2024

  4. [4]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021

    2021

  5. [5]

    Openclip, July 2021

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. URL https://doi.org/10.5281/ zenodo.5143773. If you use this software, please cite it as below

    2021

  6. [6]

    Plain-det: A plain multi-dataset object detector

    Cheng Shi, Yuchen Zhu, and Sibei Yang. Plain-det: A plain multi-dataset object detector. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part V, page 210–226, Berlin, Heidelberg, 2024. Springer-Verlag. ISBN 978-3-031-72651-4. doi: 10.1007/978-3-031-72652-1_13. URL https://doi.org/10. 1007/...

    work page doi:10.1007/978-3-031-72652-1_13 2024

  7. [7]

    LMSeg: Language-guided multi-dataset segmentation

    Qiang Zhou, Yuang Liu, Chaohui Yu, Jingliang Li, Zhibin Wang, and Fan Wang. LMSeg: Language-guided multi-dataset segmentation. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=P44WPn1_ aJV

    work page 2023

  8. [8]

    Detection hub: Unifying object detection datasets via query adaptation on language embedding

    Lingchen Meng, Xiyang Dai, Yinpeng Chen, Pengchuan Zhang, Dongdong Chen, Mengchen Liu, Jianfeng Wang, Zuxuan Wu, Lu Yuan, and Yu-Gang Jiang. Detection hub: Unifying object detection datasets via query adaptation on language embedding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11402–11411, June 2023

    work page 2023

Show all 37 references
  1. [9]

    Universal-rcnn: Universal object detector via transferable graph r-cnn

    Hang Xu, Linpu Fang, Xiaodan Liang, Wenxiong Kang, and Zhenguo Li. Universal-rcnn: Universal object detector via transferable graph r-cnn. Proceedings of the AAAI Conference on Artificial Intelligence, 34:12492–12499, 04 2020. doi: 10.1609/aaai.v34i07.6937

    2020 doi

  2. [10]

    Automated label unification for multi-dataset semantic segmentation with GNNs

    Rong Ma, Jie Chen, Xiangyang Xue, and Jian Pu. Automated label unification for multi-dataset semantic segmentation with GNNs. In The Thirty-eighth Annual Conference on Neural Informa- tion Processing Systems, 2024. URL https://openreview.net/forum?id=gSGLkCX9sc. 10

    2024

  3. [11]

    MSeg: A composite dataset for multi-domain semantic segmentation

    John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. MSeg: A composite dataset for multi-domain semantic segmentation. In Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  4. [12]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks

    Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, 2013

    2013

  5. [13]

    Unbiased teacher for semi-supervised object detection

    Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In Proceedings of the International Conference on Learning Representations (ICLR), 2021

    2021

  6. [14]

    Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli, and Robby T. Tan. 2pcnet: Two- phase consistency training for day-to-night unsupervised domain adaptive object detection. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  7. [15]

    Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli, and Robby T. Tan. Cat: Exploiting inter-class dynamics for domain adaptive object detection. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  8. [16]

    Cross-domain adaptive teacher for object detection

    Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  9. [17]

    Cycada: Cycle-consistent adversarial domain adaptation

    Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In Interna- tional conference on machine learning, 2018

    2018

  10. [18]

    Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation

    Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. arXiv preprint arXiv:1705.00609, 2017

    2017 arXiv

  11. [19]

    Forkgan: Seeing into the rainy night

    Ziqiang Zheng, Yang Wu, Xinran Han, and Jianbo Shi. Forkgan: Seeing into the rainy night. In The IEEE European Conference on Computer Vision (ECCV), August 2020

    2020

  12. [20]

    Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization

    Prithvijit Chattopadhyay*, Kartik Sarangmath*, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In IEEE/CVF International Conference in Computer Vision (ICCV), 2023

    2023

  13. [21]

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In Kamalika Chaudhuri and Ruslan Salakhutdinov, edi- tors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of M...

    2019

  14. [22]

    Hénaff, Alexander Kolesnikov, Xiaohua Zhai, and Aäron van den Oord

    Lucas Beyer, Olivier J. Hénaff, Alexander Kolesnikov, Xiaohua Zhai, and Aäron van den Oord. Are we done with imagenet?, 2020. URL https://arxiv.org/abs/2006.07159

    2020 arXiv

  15. [23]

    Re-labeling imagenet: from single to multi-labels, from global to localized labels

    Sangdoo Yun, Seong Joon Oh, Byeongho Heo, Dongyoon Han, Junsuk Choe, and Sanghyuk Chun. Re-labeling imagenet: from single to multi-labels, from global to localized labels. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  16. [24]

    Automatic universal taxonomies for multi-domain semantic segmentation

    Petra Bevandic and Sinisa Segvic. Automatic universal taxonomies for multi-domain semantic segmentation. CoRR, abs/2207.08445, 2022. doi: 10.48550/ARXIV .2207.08445. URL https: //doi.org/10.48550/arXiv.2207.08445

    2022 doi arXiv

  17. [25]

    Multi-domain semantic segmentation with overlapping labels

    Petra Bevandi´c, Marin Orši ´c, Ivan Grubiši ´c, Josip Šari ´c, and Siniša Šegvi ´c. Multi-domain semantic segmentation with overlapping labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2615–2624, January 2022

    2022

  18. [26]

    Automated detection of label errors in semantic seg- mentation datasets via deep learning and uncertainty quantification

    Matthias Rottmann and Marco Reese. Automated detection of label errors in semantic seg- mentation datasets via deep learning and uncertainty quantification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3214–3223, January 2023. 11

    2023

  19. [27]

    Simple multi-dataset detection

    Xingyi Zhou, Vladlen Koltun, and Philipp Krähenbühl. Simple multi-dataset detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7571–7580, June 2022

    2022

  20. [28]

    Faster r-cnn: Towards real-time object detection with region proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran ...

    2015

  21. [29]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rab...

    2024

  22. [30]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognitio...

    2016

  23. [31]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. arXiv preprint arXiv:1903.11027, 2019

    1903 arXiv

  24. [32]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang,...

    2020

  25. [33]

    ACDC: The adverse conditions dataset with correspondences for semantic driving scene understanding

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. ACDC: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021

    2021

  26. [34]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVP...

    2020

  27. [35]

    SHIFT: a synthetic driving dataset for continuous multi-task domain adaptation

    Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. SHIFT: a synthetic driving dataset for continuous multi-task domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ...

    2022

  28. [36]

    Detectron2

    Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019

    2019

  29. [37]

    Synscapes: A photorealistic synthetic dataset for street scene parsing, 2018

    Magnus Wrenninge and Jonas Unger. Synscapes: A photorealistic synthetic dataset for street scene parsing, 2018. URL https://arxiv.org/abs/1810.08705. 12

    2018 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.