Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Decoupled PROB: Decoupled Query Initialization Tasks and Objectness-Class Learning for Open World Object Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that decoupling query initialization by task and stopping objectness prediction after decoder layer 2 makes a transformer-based open-world detector set state-of-the-art known-object mAP on M-OWODB and S-OWODB without…

desk verdict Solid engineering extension of PROB with two simple, well-ablated modules; the main caveat is hyperparameters tuned on Task 1 test set with no validation split. read the letter →

arxiv 2507.13085 v1 pith:JAYYPXEA submitted 2025-07-17 cs.CV

Riku Inoue , Masamitsu Tsuchiya , Yuji Yasui This is my paper
classification cs.CV
keywords OpenWorldObjectDetectionProbabilisticObjectnessQueryInitializationSelectionDeformableDETRIncrementalLearningUnknownPrediction
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

Open World Object Detection asks a detector to keep finding known classes while surfacing anything unfamiliar as "unknown" and later learning those unknowns as new classes. The paper argues that the current strong PROB transformer detector is held back by two design problems: its object queries are all learnable, so early decoder layers do not focus on objects, and its objectness head keeps learning alongside the class head, pulling query embeddings in opposite directions. Decoupled PROB fixes both by splitting the 100 queries into 20 started from class-score query selection (which lock onto known objects) and 80 learnable queries (which spread over the image to catch unknowns), and by stopping objectness prediction after the second of six decoder layers. The result, per the paper, is state-of-the-art known-object mAP on both standard OWOD benchmarks from task 2 onward while roughly holding unknown-object recall, at nearly the same parameter count and FLOPs as its baseline. If true, the message is that OWOD does not need pseudo-labels for unknowns; it needs the right division of labor between the known-focused and unknown-focused parts of the decoder.

What carries the argument

The load-bearing mechanism is the pairing of Task-Decoupled Query Initialization (TDQI) and Early Termination of Objectness Prediction (ETOP). TDQI splits a fixed pool of 100 object queries: 20 initialized from encoder top-class-score regions (query selection), which concentrate reference points on known objects from the first decoder layer, and 80 learnable queries whose reference points spread evenly and drift toward unknown objects by the last layer. ETOP cuts the objectness head off after decoder layer 2 while class and box heads keep running, so the early layers learn "is this an object?" under the focused queries, and later layers learn "which class?" and refine boxes without objectness pulling the embeddings back together.

What would settle it

Run the same model on S-OWODB while tuning the query ratio and objectness stop layer on a held-out validation split; if the Task-1-tuned 20/80, layer-2 configuration is not consistently best, the reported gains are tied to the chosen settings rather than to the method's decoupling.

Watch

Extended reading notes

Core claim

The paper's central claim is that PROB's objectness and class heads conflict because objectness learning pulls all object query embeddings closer together while class learning pushes embeddings of different classes apart, and that this conflict can be removed by Early Termination of Objectness Prediction (ETOP): objectness is predicted only in the first two decoder layers, while class and bounding box are refined in all six. To make early objectness reliable, the paper adds Task-Decoupled Query Initialization (TDQI), which initializes 20 of the 100 queries by class-score query selection (these find known objects early) and 80 as learnable queries (these cover the rest of the image, including unknowns). On M-OWODB, Decoupled PROB reports current-known mAP gains over PROB of 4.5, 3.9, and 3.2 percentage points in tasks 2 through 4; on S-OWODB the corresponding gains are 9.7, 12.9, and 8.1 points, with both benchmarks showing higher "both" mAP and roughly comparable or better U-Recall in later tasks. The paper also shows that switching off either module lowers performance, and that ETOP beats the earlier DOL scheme that stops objectness even earlier, because ETOP keeps iteratively refining boxes and classes in all layers.

Load-bearing premise

The fixed hyperparameters (20 queries from query selection, 80 learnable, objectness stopped after layer 2) were chosen on Task 1 of M-OWODB's test set and then applied everywhere, so the benchmark gains rest on that configuration staying near-optimal across tasks and datasets.

Editorial extensions

If this is right

  • Objectness and class prediction can be decoupled temporally inside a single transformer decoder, so complex ways of resolving their conflict may be unnecessary.
  • TDQI is a module that can be dropped into other Deformable DETR-based OWOD models, so the known/unknown query split is a reusable recipe rather than a full-model change.
  • The 20/80 split and layer-2 cutoff keep model size and cost close to PROB (40.9M parameters, 163.4G FLOPs), so the gains are not bought with extra compute.
  • Because the query-selection queries carry known-object detection and the learnable queries carry unknown-object detection, the model can specialize each query type's role, which is visible in reference-point visualizations.
  • On incremental object detection (PASCAL VOC 10+10, 15+5, 19+1), the same model also improves or matches PROB's mAP, suggesting the decoupling helps continual learning as well.

Reading between the lines

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

  • Beyond the paper: because the 20/80 ratio and stop-layer were tuned on Task 1's test set and fixed, the reported margin might shrink if those hyperparameters were re-tuned per task on a held-out split; a validation study would settle this.
  • Beyond the paper: the same task-decoupled split could be extended to other open-set heads, for example by giving a few "known-anchor" queries and many "coverage" queries to segmentation or tracking models, though the paper only demonstrates detection.
  • Beyond the paper: the paper's role analysis (query-selection queries account for roughly 70% of known detections, learnable queries roughly 70% of unknown detections) suggests an adaptive allocation could be learned from the data rather than fixed at 20/80, which would be a direct testable next step.
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

3 major / 5 minor

Summary. This paper presents Decoupled PROB, an open-world object detection model built on PROB. It introduces Task-Decoupled Query Initialization (TDQI), which combines 20 query-selection-initialized queries with 80 learnable queries, and Early Termination of Objectness Prediction (ETOP), which stops objectness estimation after the second decoder layer while continuing class and box refinement. The authors report state-of-the-art known-object mAP on M-OWODB and S-OWODB, and improved incremental object detection on PASCAL VOC, compared with PROB and earlier methods.

Significance. The two proposed modules are simple and well-motivated by the analysis of learning conflicts between objectness and class predictions. The ablation study (Table 7) and the comparison with DOL (Table 8) provide qualitative support for the roles of TDQI and ETOP, and TDQI is plausibly portable to other DETR-based OWOD detectors. If the evaluation were fully validated, the method would constitute a useful, low-cost improvement over PROB. However, the hyperparameter selection procedure, the inconsistency in the USD comparison, and the overstatement relative to the paper's own results limit the strength of the significance claims as presented.

major comments (3)
  1. [Tables 2-4] The 20/80 query-selection-to-learnable-query ratio and the objectness stop layer are selected by maximizing U-Recall and mAP on the M-OWODB Task 1 test set (Tables 2 and 3), and the same Task 1 numbers are reported in the main SOTA table (Table 4). Because the selection is done on the test set, the Task 1 result is optimistically biased, and the fixed configuration is then assumed to be near-optimal for Tasks 2-4 and for S-OWODB without any held-out validation or seed averaging. The spread across nearby choices is small (mAP 58.8-59.9 in Table 2; mAP 58.0-60.4 in Table 3), whereas the claimed gains over PROB on later tasks are several mAP points (e.g., +4.5, +3.9, +3.2 on M-OWODB and +9.7, +12.9, +8.1 on S-OWODB current-known mAP). This makes it difficult to attribute the reported improvements to the architectural contributions rather than to the particular hyperparameter selection. A validation split or a sensitivity analysis across configurations, ideally with multiple seeds, is needed to support the SOTA claim.
  2. [Sec. 5.3 vs Table 4] The text in Sec. 5.3 states that the authors compare USD 'without ASF' for fairness, yet Table 4 lists the baseline as 'USD-ASF' in both benchmarks. If the table indeed reports the ASF-augmented numbers, the comparison is not the fair one described; if it reports the non-ASF variant, the row label is misleading. This inconsistency matters because USD-ASF achieves the highest U-Recall on several tasks, so the interpretation of the SOTA claim depends on which variant is actually reported.
  3. [Abstract and Sec. 12] The abstract claims that Decoupled PROB 'surpasses all existing methods across several metrics,' which is overstated relative to the paper's own results. The supplementary discussion (Sec. 12) explicitly acknowledges that U-Recall is often below CAT and USD-ASF on both benchmarks (e.g., M-OWODB Task 1 U-Recall 20.3 vs CAT 23.7 and USD-ASF 21.6 in Table 4). The contribution should be framed as improved known-object mAP with roughly comparable unknown-object recall, not broad superiority.
minor comments (5)
  1. [Fig. 4 caption] The caption reads 'set to n = 2based on experimental results' - a space is missing between '2' and 'based'. Also, the full term 'Task-Decoupled Query Initialization' should be introduced before using the acronym TDQI in the abstract.
  2. [Sec. 5.3] In the paragraph on M-OWODB, 'Current known mA' should read 'Current known mAP'.
  3. [Table 6] The training hyperparameter table is in the supplementary material, but the main text never references it; readers may miss that the learning-rate drop schedule differs between M-OWODB and S-OWODB.
  4. [Sec. 12] The discussion of convergence (e.g., USD training for 41 epochs and CAT for 45 vs. the 26/10 epoch schedule in Table 6) suggests that the comparison may be affected by different training budgets; a sentence in the main text indicating the training epochs of compared methods would help.
  5. [General] The reference list has a formatting issue in reference [13] where the title is broken across lines ('Segment any- thing'). The paper also uses 'USD' and 'USD-ASF' inconsistently in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: TDQI and ETOP are architectural changes benchmarked against PROB; hyperparameter tuning on Task 1 test set is a methodological concern, not a definitional reduction.

full rationale

Decoupled PROB's claimed contribution is an architecture: TDQI (mix of 20 query-selection and 80 learnable queries) and ETOP (stop objectness after decoder layer 2), both inherited onto the PROB baseline. The paper does not derive any predicted quantity from these choices by construction; objectness remains PROB's Gaussian/Mahalanobis model (Eq. 2), and the reported mAP/U-Recall values are measured on benchmark test sets. The selection of the 20/80 split and stop-layer 2 is done empirically on M-OWODB Task 1 (Tables 2 and 3) and then fixed; this is test-set tuning or possible inflated gains, but it is not circularity under the definitions used here because the headline metrics are not definitionally equal to those hyperparameters. PROB and USD are external prior works, not self-citations, and no uniqueness theorem or ansatz is imported from the authors' own prior work. The paper's own Sec. 12 concedes U-Recall is often below CAT and USD-ASF, but this is an honesty caveat, not a circular step. Nothing in the derivation chain reduces to its own inputs, so score 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim (improved OWOD performance) rests on two hand-tuned architectural hyperparameters and four domain assumptions inherited from prior work; no new physical or conceptual entities are postulated.

free parameters (2)
  • query selection to learnable query ratio = 20/80 (out of 100 queries)
    Selected via grid search (Table 2 of paper) on M-OWODB Task 1 test set; chosen as the configuration with the highest combined U-Recall and mAP.
  • objectness stop layer = 2 (out of 6 decoder layers)
    Selected via grid search (Table 3) on the same Task 1 test set; objectness prediction is applied in layers 1-2 and stopped after layer 2.
assumptions (4)
  • domain assumption Probabilistic objectness model (Gaussian in query embedding space) from PROB remains valid
    The paper builds on PROB's Eq. (2) but does not re-derive or justify it; the central improvement rests on this inherited model.
  • domain assumption Jointly learning objectness and class in all decoder layers causes a learning conflict
    ETOP's motivation rests on the claim from USD [11] and Figure 3 that objectness (pulling object features together) and class (pushing classes apart) conflict; the paper does not measure the conflict directly.
  • domain assumption Query selection places reference points on known objects and learnable queries cover unknown objects
    TDQI's design assumes the class-score-based selection focuses on known classes, so learnable queries must cover the rest; supported only by qualitative visualizations (Fig. 2) and detection-ratio analysis (Fig. 6).
  • ad hoc to paper The 20/80 split and stop layer 2 selected on Task 1 generalize to all tasks and datasets
    The hyperparameters were tuned on M-OWODB Task 1 and then fixed for S-OWODB and all later tasks; no validation with a separate held-out split is performed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoupled PROB: Decoupled Query Initialization Tasks and Objectness-Class Learning for Open World Object Detection." pith.science (2026). https://pith.science/paper/JAYYPXEA

@misc{pith2026250713085,
  author       = {Pith},
  title        = {Pith review of: Decoupled PROB: Decoupled Query Initialization Tasks and Objectness-Class Learning for Open World Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JAYYPXEA}},
  note         = {Machine review of arXiv:2507.13085}
}
read the original abstract

Open World Object Detection (OWOD) is a challenging computer vision task that extends standard object detection by (1) detecting and classifying unknown objects without supervision, and (2) incrementally learning new object classes without forgetting previously learned ones. The absence of ground truths for unknown objects makes OWOD tasks particularly challenging. Many methods have addressed this by using pseudo-labels for unknown objects. The recently proposed Probabilistic Objectness transformer-based open-world detector (PROB) is a state-of-the-art model that does not require pseudo-labels for unknown objects, as it predicts probabilistic objectness. However, this method faces issues with learning conflicts between objectness and class predictions. To address this issue and further enhance performance, we propose a novel model, Decoupled PROB. Decoupled PROB introduces Early Termination of Objectness Prediction (ETOP) to stop objectness predictions at appropriate layers in the decoder, resolving the learning conflicts between class and objectness predictions in PROB. Additionally, we introduce Task-Decoupled Query Initialization (TDQI), which efficiently extracts features of known and unknown objects, thereby improving performance. TDQI is a query initialization method that combines query selection and learnable queries, and it is a module that can be easily integrated into existing DETR-based OWOD models. Extensive experiments on OWOD benchmarks demonstrate that Decoupled PROB surpasses all existing methods across several metrics, significantly improving performance.

Figures

Figures reproduced from arXiv: 2507.13085 by the authors.

Figure 1
Figure 1. Visualization of reference points in the proposed model. (a) is the ground truth image, where blue bounding boxes indicate known objects and red bounding boxes indicate unknown objects. (b) and (c) show the reference points in the initial and last layers of the proposed model, respectively. Blue reference points are initialized by query selection, while red reference points are initialized by learnable queries. In t… view at source ↗
Figure 2
Figure 2. Visualization of reference points and detection results for learnable query and query selection models. In (a), (c), (e), blue bounding boxes indicate known objects, and red bounding boxes indicate unknown objects. In the model using object queries initialized by query selection, the reference points are distributed around known objects and do not cover unknown objects, making it difficult to detect unknown objects.… view at source ↗
Figure 3
Figure 3. Learning of objectness and class classification. Ob￾jectness aims to separate objects from the background by bringing the features of objects closer together, whereas class classifica￾tion aims to separate the features among different object classes. Therefore, these processes work in opposite directions [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overview of our proposed model. Our model processes the input image with a backbone and a deformable transformer encoder, same as PROB. Then, query selection is performed on the encoder features, which are concatenated with learnable queries to form the object queries.…
Figure 5
Figure 5. Figure 5: Qualitative results on selected images from the test set. The top row shows the detection results of Decoupled PROB, while the bottom row shows the detection results of PROB. Green bounding boxes indicate known classes, and yellow bounding boxes indicate unknown classe…
Figure 6
Figure 6. Figure 6: Comparison of known class detection and unknown class detection roles in each task for TDQI. initialized with learnable query. Although object queries initialized with query selection account for only about 20% of the total, they are responsi￾ble for nearly 70% of the …
Figure 7
Figure 7. Figure 7: Learning Unknown Classes as Known Classes (Incremental Learning). PROB [44], which we used as our baseline, is a remark￾able approach. However, it may have an issue where back￾ground information is included in the query embeddings used to update the objectness distribu…
Figure 8
Figure 8. Figure 8: Qualitative results on example images from test set. thing. In Proceedings of the IEEE/CVF International Con￾ference on Computer Vision, pages 4015–4026, 2023. 3, 7 [14] Nishant Kumar, Sinisaˇ Segvi ˇ c, Abouzar Eslami, and Ste- ´ fan Gumhold. Normalizing flow based fe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 35 canonical work pages

  1. [1]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European confer- ence on computer vision, pages 213–229. Springer, 2020. 1, 3

    work page 2020

  2. [2]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 6

    2021

  3. [3]

    Learning open set network with discriminative reciprocal points

    Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 507–522. Springer, 2020. 2

    2020

  4. [4]

    Group detr: Fast detr training with group-wise one-to-many assignment

    Qiang Chen, Xiaokang Chen, Jian Wang, Shan Zhang, Kun Yao, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise one-to-many assignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 6633–6642, 2023. 1

    work page 2023

  5. [5]

    Dynamic detr: End-to-end object detection with dynamic attention

    Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988–2997, 2021. 3

    work page 2021

  6. [6]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. InInternational Con- ference on Learning Representations, 2020. 3

    work page 2020

  7. [7]

    Unknown-aware object detection: Learning what you don’t know from videos in the wild

    Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware object detection: Learning what you don’t know from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13678–13688, 2022. 1

    work page 2022

  8. [8]

    The pascal visual object classes (voc) challenge

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303–338, 2010. 6

    2010

Show all 44 references
  1. [9]

    Ow-detr: Open-world detection transformer

    Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9235–9244, 2022. 1, 2, 6, 7, 8, 10, 11

    2022

  2. [10]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 6

    2016

  3. [11]

    Usd: Unknown sensitive detector empowered by decoupled ob- jectness and segment anything model

    Yulin He, Wei Chen, Yusong Tan, and Siqi Wang. Usd: Unknown sensitive detector empowered by decoupled ob- jectness and segment anything model. arXiv preprint arXiv:2306.02275, 2023. 2, 3, 4, 5, 6, 7, 9

    2023 arXiv

  4. [12]

    Towards open world object de- tection

    KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vi- neeth N Balasubramanian. Towards open world object de- tection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5830–5840,

  5. [13]

    Segment any- 12 Figure 8

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- 12 Figure 8. Qualitative results on example images from test set. thing. In Proceedings of the IEEE/CVF I...

    2023

  6. [14]

    Normalizing flow based feature synthe- sis for outlier-aware object detection

    Nishant Kumar, Sini ˇsa ˇSegvi´c, Abouzar Eslami, and Ste- fan Gumhold. Normalizing flow based feature synthe- sis for outlier-aware object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5156–5165, 2023. 1

    2023

  7. [15]

    Lite detr: An interleaved multi-scale encoder for efficient detr

    Feng Li, Ailing Zeng, Shilong Liu, Hao Zhang, Hongyang Li, Lei Zhang, and Lionel M Ni. Lite detr: An interleaved multi-scale encoder for efficient detr. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18558–18567, 2023. 1

    2023

  8. [16]

    Dn-detr: Accelerate detr training by intro- 13 ducing query denoising

    Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by intro- 13 ducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 13619–13627, 2022. 3

    2022

  9. [17]

    Unknown sniffer for object detection: Don’t turn a blind eye to unknown objects

    Wenteng Liang, Feng Xue, Yihao Liu, Guofeng Zhong, and Anlong Ming. Unknown sniffer for object detection: Don’t turn a blind eye to unknown objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3230–3239, 2023. 1

    2023

  10. [18]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceeding...

    2014

  11. [19]

    Dab-detr: Dynamic an- chor boxes are better queries for detr

    Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic an- chor boxes are better queries for detr. In International Con- ference on Learning Representations, 2021. 3

    2021

  12. [20]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. In International Conference on Learning Representations, 2018. 6

    2018

  13. [21]

    Cat: Localization and iden- tification cascade detection transformer for open-world ob- ject detection

    Shuailei Ma, Yuefeng Wang, Ying Wei, Jiaqi Fan, Thomas H Li, Hongli Liu, and Fanbing Lv. Cat: Localization and iden- tification cascade detection transformer for open-world ob- ject detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2023

  14. [22]

    Rethinking open-world object detection in autonomous driving scenarios

    Zeyu Ma, Yang Yang, Guoqing Wang, Xing Xu, Heng Tao Shen, and Mingxing Zhang. Rethinking open-world object detection in autonomous driving scenarios. In Proceedings of the 30th ACM International Conference on Multimedia , pages 1279–1288, 2022. 1

    2022

  15. [23]

    Conditional detr for fast training convergence

    Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceed- ings of the IEEE/CVF international conference on computer vision, pages 3651–3660, 2021. 3

    2021

  16. [24]

    Faster ilod: In- cremental learning for object detectors based on faster rcnn

    Can Peng, Kun Zhao, and Brian C Lovell. Faster ilod: In- cremental learning for object detectors based on faster rcnn. Pattern recognition letters, 140:109–115, 2020. 8

    2020

  17. [25]

    Rank-detr for high quality object detection

    Yifan Pu, Weicong Liang, Yiduo Hao, Yuhui Yuan, Yukang Yang, Chao Zhang, Han Hu, and Gao Huang. Rank-detr for high quality object detection. Advances in Neural Informa- tion Processing Systems, 36, 2024. 1, 3

    2024

  18. [26]

    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. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137–1149, 2016. 2

    2016

  19. [27]

    Sparse detr: Efficient end-to-end object detection with learnable sparsity

    Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Sae- hoon Kim. Sparse detr: Efficient end-to-end object detection with learnable sparsity. arXiv preprint arXiv:2111.14330 ,

    arXiv

  20. [28]

    Learning to detect every thing in an open world

    Kuniaki Saito, Ping Hu, Trevor Darrell, and Kate Saenko. Learning to detect every thing in an open world. InEuropean Conference on Computer Vision , pages 268–284. Springer,

  21. [29]

    Incremental learning of object detectors without catas- trophic forgetting

    Konstantin Shmelkov, Cordelia Schmid, and Karteek Ala- hari. Incremental learning of object detectors without catas- trophic forgetting. In Proceedings of the IEEE international conference on computer vision, pages 3400–3409, 2017. 8

    2017

  22. [30]

    Order: Open world object detection on road scenes

    Deepak Kumar Singh, Shyam Nandan Rai, KJ Joseph, Rohit Saluja, Vineeth N Balasubramanian, Chetan Arora, Anbu- mani Subramanian, and CV Jawahar. Order: Open world object detection on road scenes. In Proc. NeurIPS Work- shops, volume 1, page 3, 2021. 1

    2021

  23. [31]

    Exploring or- thogonality in open world object detection

    Zhicheng Sun, Jinghan Li, and Yadong Mu. Exploring or- thogonality in open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 17302–17312, 2024. 11

    2024

  24. [32]

    Selective search for ob- ject recognition

    Jasper RR Uijlings, Koen EA Van De Sande, Theo Gev- ers, and Arnold WM Smeulders. Selective search for ob- ject recognition. International journal of computer vision , 104:154–171, 2013. 3

    2013

  25. [33]

    Anchor detr: Query design for transformer-based detector

    Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelli- gence, volume 36, pages 2567–2575, 2022. 3

    2022

  26. [34]

    Discriminating known from unknown objects via structure-enhanced recurrent varia- tional autoencoder

    Aming Wu and Cheng Deng. Discriminating known from unknown objects via structure-enhanced recurrent varia- tional autoencoder. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 23956–23965, 2023. 1

    2023

  27. [35]

    Two-branch objectness-centric open world de- tection

    Yan Wu, Xiaowei Zhao, Yuqing Ma, Duorui Wang, and Xi- anglong Liu. Two-branch objectness-centric open world de- tection. In Proceedings of the 3rd International Workshop on Human-Centric Multimedia Analysis, pages 35–40, 2022. 2, 7

    2022

  28. [36]

    Uc-owod: Unknown-classified open world object detection

    Zhiheng Wu, Yue Lu, Xingyu Chen, Zhengxing Wu, Liwen Kang, and Junzhi Yu. Uc-owod: Unknown-classified open world object detection. In European Conference on Com- puter Vision, pages 193–210. Springer, 2022. 2, 7

    2022

  29. [37]

    Convolutional prototype network for open set recognition

    Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, Qing Yang, and Cheng-Lin Liu. Convolutional prototype network for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2358–2370, 2020. 2

    2020

  30. [38]

    Efficient detr: improving end-to-end object detector with dense prior

    Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 2, 3, 5, 9

    2021 arXiv

  31. [39]

    Open-world object detection via discriminative class prototype learning

    Jinan Yu, Liyan Ma, Zhenglin Li, Yan Peng, and Shaorong Xie. Open-world object detection via discriminative class prototype learning. In 2022 IEEE International Conference on Image Processing (ICIP), pages 626–630. IEEE, 2022. 2, 7

    2022

  32. [40]

    Dino: Detr with improved denoising anchor boxes for end-to-end object de- tection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object de- tection. In The Eleventh International Conference on Learn- ing Representations, 2022. 1, 3

    2022

  33. [41]

    Revisiting open world object detection

    Xiaowei Zhao, Yuqing Ma, Duorui Wang, Yifan Shen, Yix- uan Qiao, and Xianglong Liu. Revisiting open world object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1

    2023

  34. [42]

    14 Detrs beat yolos on real-time object detection.arXiv preprint arXiv:2304.08069, 2023

    Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. 14 Detrs beat yolos on real-time object detection.arXiv preprint arXiv:2304.08069, 2023. 1, 3

    2023 arXiv

  35. [43]

    Deformable detr: Deformable transformers for end-to-end object detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 2, 3, 5, 6, 9, 10, 11

    2021

  36. [44]

    Prob: Probabilistic objectness for open world object detection

    Orr Zohar, Kuan-Chieh Wang, and Serena Yeung. Prob: Probabilistic objectness for open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 11444–11453, 2023. 2, 3, 4, 6, 7, 8, 9, 11, 12 15

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.