Back

PROB: Probabilistic Objectness for Open World Object Detection

Orr ZoharKuan-Chieh WangSerena Yeung

article2023CVPR133 citations

Presents a probabilistic framework that models feature-space objectness distributions to distinguish unknown objects from background without pseudo-labels, doubling unknown object recall over prior open-world detection methods.

Listen

Standard computer vision models struggle when deployed in real-world environments because they assume all possible objects belong to fixed, predefined categories. When traditional detectors encounter novel or unlabeled items, they incorrectly classify them as background. Open-world object detection aims to bridge this gap by enabling models to recognize known items, flag novel objects for human annotation, and incrementally learn these new classes without forgetting earlier knowledge. However, existing methods suffer from extremely low unknown object recall (typically around 10%) because distinguishing unlabeled objects from true background without direct supervision remains fundamentally difficult.

The main objective of the article is to introduce and evaluate PROB (Probabilistic Objectness Open World Detection Transformer), a framework that estimates general objectness via a probabilistic model to improve both unknown object discovery and continuous learning.

The authors implemented PROB by integrating a probabilistic objectness head into a transformer-based object detection architecture. Rather than relying on heuristic pseudo-labels to guess unknown objects during training, the approach models the distribution of query features as a class-agnostic Gaussian distribution. Training alternates between estimating this probability distribution and maximizing the likelihood of known objects. For incremental updates, the system selects stored representative images (exemplars) based on high and low objectness scores. The method was evaluated on standard open-world benchmarks built on MS-COCO and PASCAL VOC datasets across multiple incremental tasks, tracking unknown recall, classification error, and known mean average precision.

The evaluation yielded several key findings. First, PROB doubled to tripled unknown object recall across all tasks, achieving relative gains of 100% to 300% over prior state-of-the-art methods (for example, reaching 17.6% to 24.8% unknown recall on the separated-superclass benchmark compared to 5.7% to 6.9% for the leading baseline). Second, the method significantly reduced the rate of unknown objects being mistakenly classified as known classes, lowering open-set errors by roughly 25% to 60%. Third, PROB improved known object detection accuracy by approximately 10% across tasks and maintained higher accuracy during incremental updates, demonstrating superior resistance to catastrophic forgetting.

These findings indicate that decoupling general objectness estimation from specific category classification resolves the core tension between identifying novel objects and ignoring background. For operational computer vision systems in robotics, autonomous driving, and healthcare, this reduces the safety risks associated with unflagged novel hazards while optimizing human-in-the-loop annotation workflows by delivering more reliable candidate detections.

Organizations developing open-world computer vision systems should consider adopting probabilistic density estimation over heuristic pseudo-labeling. Practitioners should also leverage objectness-driven exemplar selection to maintain performance during continuous retraining. Before broad deployment, engineering teams should conduct domain-specific pilot studies on custom video feeds or specialized sensors to establish baseline operational confidence.

The primary limitation of this work is that overall unknown object recall remains below 25%, meaning most novel objects still go undetected despite the substantial relative gains. Additionally, the evaluations rely on standardized benchmark datasets, which may not capture all operational complexities of open-ended physical environments. Consequently, while the probabilistic mechanism is sound and outperforms current alternatives, users should maintain human oversight until detection rates on novel classes improve further.

Stanford Universitycomputer visiondetection algorithmencoder-decoderopen world object detectionprobabilistic objectnessclass-agnostic object detection
arXiv: 2212.01424orrzohar/PROB
Cover for PROB: Probabilistic Objectness for Open World Object Detection

Abstract

Open World Object Detection (OWOD) is a new and challenging computer vision task that bridges the gap between classic object detection (OD) benchmarks and object detection in the real world. In addition to detecting and classifying seen/labeled objects, OWOD algorithms are expected to detect novel/unknown objects - which can be classified and incrementally learned. In standard OD, object proposals not overlapping with a labeled object are automatically classified as background. Therefore, simply applying OD methods to OWOD fails as unknown objects would be predicted as background. The challenge of detecting unknown objects stems from the lack of supervision in distinguishing unknown objects and background object proposals. Previous OWOD methods have attempted to overcome this issue by generating supervision using pseudo-labeling - however, unknown object detection has remained low. Probabilistic/generative models may provide a solution for this challenge. Herein, we introduce a novel probabilistic framework for objectness estimation, where we alternate between probability distribution estimation and objectness likelihood maximization of known objects in the embedded feature space - ultimately allowing us to estimate the objectness probability of different proposals. The resulting Probabilistic Objectness transformer-based open-world detector, PROB, integrates our framework into traditional object detection models, adapting them for the open-world setting. Comprehensive experiments on OWOD benchmarks show that PROB outperforms all existing OWOD methods in both unknown object detection (~ 2× unknown recall) and known object detection (~ 10% mAP). Our code is available at https://github.com/orrzohar/PROB.

Topics

computer visiondetection algorithmencoder-decoderfeature spacegenerative modelprobability distributionsclassifiersestimation methodobject detection algorithms

Keywords

open world object detectionprobabilistic objectnessclass-agnostic object detectionobjectness estimationobjectness likelihood maximizationobject proposalsunknown objectsobject recall

Table of Contents

  • 1. Introduction
  • 2. Related Works
  • 3. Background
  • 4. Method
  • 4.1. Probabilistic Objectness
  • 4.2. Objectness for Incremental Learning
  • 5. Experiments & Results
  • 5.1. Open World Object Detection Performance
  • 5.2. Ablation Study
  • 5.3. Incremental Object Detection
  • 6. Conclusions
  • References

Knowls

  1. Knowl 1 — Probabilistic Objectness Formulation and Factored Class Prediction

    model/method

    In the Probabilistic Objectness Open World Detection Transformer (PROB), object detection in an open-world setting is formulated by explicitly decoupling the objectness indicator o∈{0,1}o \in \{0, 1\} from the class label prediction l∈{0,1}Kt+1l \in \{0, 1\}^{K_t+1} conditioned on objectness, where KtK_t denotes the number of known classes at task tt and the first dimension represents unknown objects. Assuming background proposals satisfy p(l∣o=0,q)=0p(l \mid o = 0, q) = 0, the conditional class distribution given a query embedding q∈RDq \in \mathbb{R}^D factors as:

    p(l∣q)=∑i∈{0,1}p(l∣o=i,q)⋅p(o=i∣q)=p(l∣o=1,q)⋅p(o=1∣q)p(l \mid q) = \sum_{i \in \{0, 1\}} p(l \mid o = i, q) \cdot p(o = i \mid q) = p(l \mid o = 1, q) \cdot p(o = 1 \mid q)

    The classification head fclst(q)f_{\text{cls}}^t(q) is trained to approximate p(l∣o=1,q)p(l \mid o = 1, q), operating under the assumption that the proposal represents an object. The objectness probability p(o=1∣q)p(o = 1 \mid q) is parameterized as a class-agnostic multivariate Gaussian distribution in query embedding space, o∣q∼N(μ,Σ)o \mid q \sim \mathcal{N}(\mu, \Sigma), where μ∈RD\mu \in \mathbb{R}^D is the mean vector and Σ∈RD×D\Sigma \in \mathbb{R}^{D \times D} is the covariance matrix. The objectness likelihood fobjt(q)∈(0,1]f_{\text{obj}}^t(q) \in (0, 1] is computed using the Mahalanobis distance dM(q)d_M(q):

    fobjt(q)=exp⁡(−(q−μ)TΣ−1(q−μ))=exp⁡(−dM(q)2)f_{\text{obj}}^t(q) = \exp\left(-(q - \mu)^T \Sigma^{-1} (q - \mu)\right) = \exp\left(-d_M(q)^2\right)

    The final class prediction vector is the element-wise product:

    p(l∣q)=fclst(q)⋅fobjt(q)p(l \mid q) = f_{\text{cls}}^t(q) \cdot f_{\text{obj}}^t(q)

    When a query embedding represents background, fobjt(q)≈0f_{\text{obj}}^t(q) \approx 0, suppressing false object predictions without requiring explicit negative background-versus-unknown supervision.

  2. Knowl 2 — Two-Step Alternating Optimization for Objectness Distribution and Likelihood

    algorithm

    PROB trains its probabilistic objectness head via an alternating two-step optimization procedure without requiring explicit unknown-object pseudo-labeling:

    Input: Set of all query embeddings Q={q1,…,qNquery}⊂RDQ = \{q_1, \dots, q_{N_{\text{query}}}\} \subset \mathbb{R}^D from mini-batches, matched query indices Z\mathcal{Z} from Hungarian matching, current distribution parameters (μ,Σ)(\mu, \Sigma), EMA momentum α=0.1\alpha = 0.1.
    Output: Updated distribution parameters (μ,Σ)(\mu, \Sigma) and objectness loss Lo\mathcal{L}_o.
    Step 1: Distribution Estimation
      Compute empirical batch mean μ^=1∣Q∣∑q∈Qq\hat{\mu} = \frac{1}{|Q|} \sum_{q \in Q} q
      Compute empirical batch covariance Σ^=1∣Q∣∑q∈Q(q−μ^)(q−μ^)T\hat{\Sigma} = \frac{1}{|Q|} \sum_{q \in Q} (q - \hat{\mu})(q - \hat{\mu})^T
      Update global mean: μ←(1−α)μ+αμ^\mu \leftarrow (1 - \alpha)\mu + \alpha \hat{\mu}
      Update global covariance: Σ←(1−α)Σ+αΣ^\Sigma \leftarrow (1 - \alpha)\Sigma + \alpha \hat{\Sigma}
    Step 2: Likelihood Maximization
      Compute squared Mahalanobis distance for each matched query i∈Zi \in \mathcal{Z}:
        dM(qi)2=(qi−μ)TΣ−1(qi−μ)d_M(q_i)^2 = (q_i - \mu)^T \Sigma^{-1} (q_i - \mu)
      Compute objectness loss Lo=∑i∈ZdM(qi)2\mathcal{L}_o = \sum_{i \in \mathcal{Z}} d_M(q_i)^2
      Backpropagate Lo\mathcal{L}_o to optimize the feature extractor and decoder

    Penalizing Lo\mathcal{L}_o forces matched query embeddings corresponding to ground-truth objects to migrate toward the distribution center μ\mu, increasing their objectness likelihood fobjt(q)=exp⁡(−dM(q)2)f_{\text{obj}}^t(q) = \exp(-d_M(q)^2), while unmatched background queries remain unpenalized and receive low likelihood values.

  3. Knowl 3 — Extreme-Objectness Active Exemplar Selection for Incremental Learning

    model/method

    To mitigate catastrophic forgetting across successive tasks t∈{1,…,T}t \in \{1, \dots, T\} in Open World Object Detection (OWOD), PROB stores a set of exemplars from previous tasks using an active selection criterion based on objectness scores rather than uniform random sampling.

    After training on task dataset Dt\mathcal{D}^t, the objectness likelihood fobjt(qi)=exp⁡(−dM(qi)2)f_{\text{obj}}^t(q_i) = \exp(-d_M(q_i)^2) is evaluated for every query embedding qiq_i matched to a ground-truth object instance. For each object class:

    1. The top 25 instances with the highest objectness scores are selected as exemplars. These instances are highly representative, prototypical samples of the class that help prevent catastrophic forgetting of previously acquired representations.
    2. The bottom 25 instances with the lowest objectness scores are selected as exemplars. These instances represent hard, ambiguous cases near the objectness decision boundary, and retraining on them improves model robustness and discrimination for newly introduced classes.

    If the total number of selected exemplars exceeds the predetermined storage memory budget, the pool is randomly subsampled down to the fixed budget size.

  4. Knowl 4 — PROB Base Architecture and Multi-Task Loss

    model/method

    PROB is built upon Deformable DETR (D-DETR) with a ResNet-50 backbone initialized with self-supervised DINO pre-training and a Feature Pyramid Network (FPN). The deformable transformer encoder-decoder transforms multi-scale feature maps into Nquery=100N_{\text{query}} = 100 learned query embeddings q∈RDq \in \mathbb{R}^D with feature dimensionality D=256D = 256.

    The query embeddings are processed by three parallel heads:

    1. A bounding box regression head fbboxtf_{\text{bbox}}^t trained using smooth L1L_1 loss and generalized Intersection-over-Union (gIoU) loss, denoted jointly as Lb\mathcal{L}_b.
    2. A classification head fclstf_{\text{cls}}^t that outputs class probability distributions across known classes and the unknown class, trained via sigmoid focal loss Lc\mathcal{L}_c.
    3. A probabilistic objectness head fobjtf_{\text{obj}}^t parameterized by mean μ\mu and covariance Σ\Sigma, trained via Mahalanobis distance penalization Lo=∑i∈ZdM(qi)2\mathcal{L}_o = \sum_{i \in \mathcal{Z}} d_M(q_i)^2 on matched query indices Z\mathcal{Z}.

    Bipartite matching between ground-truth objects and query predictions is established using the Hungarian matching algorithm prior to loss calculation.

  5. Knowl 5 — Open World Object Detection Performance on M-OWODB and S-OWODB

    data/table

    Performance of PROB compared against state-of-the-art open-world object detectors on the superclass-mixed benchmark (M-OWODB) and superclass-separated benchmark (S-OWODB). Metrics reported are Unknown Recall (U-Recall, %) and known class mean Average Precision at IoU 0.5 ([email protected], %) split into Previously Known, Current Known, and Both.

    Task Task 1 Task 2 Task 3 Task 4
    Metric U-Rec Curr U-Rec Prev Curr Both U-Rec Prev Curr Both Prev Curr Both
    M-OWODB Benchmark
    ORE* 4.9 56.0 2.9 52.7 26.0 39.4 3.9 38.2 12.7 29.7 29.6 12.4 25.3
    UC-OWOD 2.4 50.7 3.4 33.1 30.5 31.8 8.7 28.8 16.3 24.6 25.6 15.9 23.2
    OCPL 8.26 56.6 7.65 50.6 27.5 39.1 11.9 38.7 14.7 30.7 30.7 14.4 26.7
    2B-OCD 12.1 56.4 9.4 51.6 25.3 38.5 11.6 37.2 13.2 29.2 30.0 13.3 25.8
    OW-DETR 7.5 59.2 6.2 53.6 33.5 42.9 5.7 38.3 15.8 30.8 31.4 17.1 27.8
    PROB (Ours) 19.4 59.5 17.4 55.7 32.2 44.0 19.6 43.0 22.2 36.0 35.7 18.9 31.5
    S-OWODB Benchmark
    ORE* 1.5 61.4 3.9 56.5 26.1 40.6 3.6 38.7 23.7 33.7 33.6 26.3 31.8
    OW-DETR 5.7 71.5 6.2 62.8 27.5 43.8 6.9 45.2 24.9 38.5 38.2 28.1 33.1
    PROB (Ours) 17.6 73.4 22.3 66.3 36.0 50.4 24.8 47.8 30.4 42.0 42.6 31.7 39.9

    PROB achieves a 2- to 3-fold improvement in Unknown Object Recall across all tasks on both benchmarks (e.g., 19.4% vs 7.5% on Task 1 M-OWODB and 17.6% vs 5.7% on Task 1 S-OWODB compared to OW-DETR). Furthermore, PROB improves known class mAP across all incremental tasks (reaching 31.5% and 39.9% in Task 4 on M-OWODB and S-OWODB, respectively) and exhibits lower drops in Previously Known mAP between tasks.

  6. Knowl 6 — Ablation of PROB Components on M-OWODB

    data/table

    Ablation study evaluating the contribution of each architectural and training component of PROB on the M-OWODB benchmark. Evaluated variants include: PROB-Obj (disables objectness likelihood maximization Lo\mathcal{L}_o, retaining only distribution parameter estimation), PROB-L2 (replaces Mahalanobis distance with Euclidean L2L_2 loss, equivalent to assuming μ=0,Σ=I\mu = 0, \Sigma = I), and PROB-IL (disables active extreme-objectness exemplar selection, using random selection).

    Task Task 1 Task 2 Task 3 Task 4
    Model U-Rec Curr U-Rec Prev Curr Both U-Rec Prev Curr Both Prev Curr Both
    Upper Bound 31.6 62.5 40.5 55.8 38.1 46.9 42.6 42.4 29.3 33.9 35.6 23.1 32.5
    D-DETR - 60.3 - 54.5 34.4 44.7 - 40.0 17.7 33.3 32.5 20.0 29.4
    PROB-Obj 21.1 39.3 18.9 41.0 23.5 32.3 22.2 34.7 16.3 28.6 29.2 13.4 25.2
    PROB-L2 22.9 53.4 19.8 49.4 28.5 39.4 21.9 37.4 15.7 30.2 30.7 14.8 26.7
    PROB-IL 19.4 59.5 15.9 54.7 32.2 43.5 18.4 42.6 20.7 35.3 34.7 17.4 30.4
    Final PROB 19.4 59.5 17.4 55.7 32.2 44.0 19.6 43.0 22.2 36.0 35.7 18.9 31.5

    Disabling likelihood maximization (PROB-Obj) causes the objectness score to uncalibrate, failing to suppress background query embeddings and degrading known object mAP (Task 1 drops from 59.5% to 39.3%). Replacing Mahalanobis distance with L2L_2 distance (PROB-L2) fails to capture query feature covariances, reducing Task 1 mAP to 53.4%. Disabling active exemplar replay (PROB-IL) reduces both U-Recall and known class retention in subsequent tasks.

  7. Knowl 7 — Incremental Object Detection Performance on PASCAL VOC 2007

    data/table

    State-of-the-art comparison on the PASCAL VOC 2007 incremental object detection (iOD) benchmark across three class splits: 10 + 10 classes, 15 + 5 classes, and 19 + 1 class(es). In each split, the detector is first trained on the base set of classes and incrementally learns the remaining classes.

    Setting / Method Old Classes (mAP) New Classes (mAP) Final mAP (%)
    10 + 10 setting
    ILOD 63.2 63.2 63.2
    Faster ILOD 69.8 54.5 62.1
    ORE −- EBUI 60.4 68.8 64.5
    OW-DETR 63.5 67.9 65.7
    PROB (Ours) 66.0 67.2 66.5
    15 + 5 setting
    ILOD 68.3 58.4 65.8
    Faster ILOD 71.6 56.9 67.9
    ORE −- EBUI 71.8 58.7 68.5
    OW-DETR 72.2 59.8 69.4
    PROB (Ours) 73.2 60.8 70.1
    19 + 1 setting
    ILOD 68.5 62.7 68.2
    Faster ILOD 68.9 61.1 68.5
    ORE −- EBUI 69.4 60.1 68.8
    OW-DETR 70.2 62.0 70.2
    PROB (Ours) 73.9 48.5 72.6

    PROB achieves the highest overall final mAP across all three incremental scenarios (66.5% for 10+10, 70.1% for 15+5, and 72.6% for 19+1), demonstrating superior retention on old classes (e.g., 73.9% vs 70.2% for OW-DETR in the 19+1 split) due to its active exemplar selection strategy.

Coverage note — None was omitted; all primary contributions, mathematical formulations, training algorithms, exemplar selection strategies, benchmark results (M-OWODB, S-OWODB, PASCAL VOC incremental OD), and ablation studies from the paper have been captured.

References

  1. 1.Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers, 2020. 3
  2. 2.Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5
  3. 3.Francisco M. Castro, Manuel J. Marin-Jimenez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 5
  4. 4.Yinong Chen and Gennaro De Luca. Technologies supporting artificial intelligence and robotics application development. Journal of Artificial Intelligence and Technology, 1(1):1–8, Jan. 2021. 1
  5. 5.Akshay Raj Dhamija, Manuel G"unther, Jonathan Ventura, and Terrance E. Boult. The overlooked elephant of object detection: Open set. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1010–1019, 2020. 1
  6. 6.R. Elakkiya, V. Subramaniyaswamy, V. Vijayakumar, and Aniket Mahanti. Cervical cancer diagnostics healthcare system using hybrid object detection adversarial networks. IEEE Journal of Biomedical and Health Informatics, 26(4):1464–1471, 2022. 1
  7. 7.Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 5
  8. 8.Akshita Gupta, Sanath Narayan, K J 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 (CVPR), pages 9235–9244, June 2022. 1, 2, 3, 5, 6, 8
  9. 9.Mohsen Jafarzadeh, Akshay Raj Dhamija, Steve Cruz, Chunchun Li, Touqeer Ahmad, and Terrance E. Boult. Open-world learning without labels. CoRR, abs/2011.12906, 2020. 1
  10. 10.K J Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5830–5840, June 2021. 1, 2, 3, 5, 6, 8
  11. 11.Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In So Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify, 2021. 2, 3, 6
  12. 12.Louis Lecrosnier, Redouane Khemmar, Nicolas Ragot, Benoit Decoux, Romain Rossi, Naceur Kefi, and Jean-Yves Ertaud. Deep learning-based object detection, localisation and tracking for smart wheelchair healthcare mobility. International Journal of Environmental Research and Public Health, 18(1), 2021. 1
  13. 13.Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018. 4
  14. 14.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 ECCV, 2014. 5
  15. 15.Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21464–21475. Curran Associates, Inc., 2020. 4
  16. 16.Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa ˇ O ˇsep, and Laura Leal-Taix'e. Opening up open world tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19045–19055, June 2022. 1
  17. 17.Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. CoRR, abs/1904.05160, 2019. 1
  18. 18.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, MM ’22, page 1279–1288, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2
  19. 19.Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multi-modal transformer. In 17th European Conference on Computer Vision (ECCV). Springer, 2022. 2, 5
  20. 20.Can Peng, Kun Zhao, and Brian C Lovell. Faster ilod: Incremental learning for object detectors based on faster rcnn. PRL, 2020. 8
  21. 21.Lu Qi, Jason Kuen, Yi Wang, Jiuxiang Gu, Hengshuang Zhao, Zhe Lin, Philip H. S. Torr, and Jiaya Jia. Open-world entity segmentation. CoRR, abs/2107.14228, 2021. 1
  22. 22.Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 5
  23. 23.Kuniaki Saito, Ping Hu, Trevor Darrell, and Kate Saenko. Learning to detect every thing in an open world, 2021. 3
  24. 24.Konstantin Shmelkov, Cordelia Schmid, and Karteek Alahari. Incremental learning of object detectors without catastrophic forgetting. In ICCV, 2017. 8
  25. 25.Deepak Kumar Singh, Shyam Nandan Rai, KJ Joseph, Rohit Saluja, Vineeth N Balasubramanian, Chetan Arora, Anbumani Subramanian, and CV Jawahar. Order: Open world object detection on road scenes. In Proc. NeurIPS Workshops, 2021. 1, 2
  26. 26.Kuan-Chieh Wang, Paul Vicol, Eleni Triantafillou, and Richard Zemel. Few-shot out-of-distribution detection. International Conference on Machine Learning (ICML) Workshop on Uncertainty and Robustness in Deep Learning, 2020. 4
  27. 27.Weiyao Wang, Matt Feiszli, Heng Wang, Jitendra Malik, and Du Tran. Open-world instance segmentation: Exploiting pseudo ground truth from learned pairwise affinity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4422–4432, June 2022. 1
  28. 28.Jiaxi Wu, Jiaxin Chen, and Di Huang. Entropy-based active learning for object detection with progressive diversity constraint. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9397–9406, June 2022. 5
  29. 29.Yan Wu, Xiaowei Zhao, Yuqing Ma, Duorui Wang, and Xianglong Liu. Two-branch objectness-centric open world detection. In Proceedings of the 3rd International Workshop on Human-Centric Multimedia Analysis, HCMA ’22, page 35–40, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2, 3, 5, 6
  30. 30.Zhiheng Wu, Yue Lu, Xingyu Chen, Zhengxing Wu, Liwen Kang, and Junzhi Yu. Uc-owod: Unknown-classified open world object detection, 2022. 1, 2, 6
  31. 31.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. 1, 2, 5, 6
  32. 32.Shengchang Zhang, Zheng Nie, and Jindong Tan. Novel objects detection for robotics grasp planning. In 2020 10th Institute of Electrical and Electronics Engineers International Conference on Cyber Technology in Automation, Control, and Intelligent Systems (CYBER), pages 43–48, 2020. 1
  33. 33.Hanbin Zhao, Hui Wang, Yongjian Fu, Fei Wu, and Xi Li. Memory-efficient class-incremental learning for image classification. IEEE Transactions on Neural Networks and Learning Systems, 33(10):5966–5977, 2022. 5
  34. 34.Xiaowei Zhao, Xianglong Liu, Yifan Shen, Yixuan Qiao, Yuqing Ma, and Duorui Wang. Revisiting open world object detection, 2022. 1, 2, 5, 6
  35. 35.Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 3, 5, 6

Citation

MLA
Zohar, O., et al. “PROB: Probabilistic Objectness for Open World Object Detection”. arXiv, 2022, http://arxiv.org/abs/2212.01424v1.
APA
Zohar, O., Wang, K.-C., & Yeung, S. (2022). PROB: Probabilistic Objectness for Open World Object Detection. arXiv. http://arxiv.org/abs/2212.01424v1
Chicago
Zohar, O., K.-C. Wang, and S. Yeung. 2022. “PROB: Probabilistic Objectness for Open World Object Detection”. arXiv. http://arxiv.org/abs/2212.01424v1.
Harvard
Zohar, O., Wang, K.-C. and Yeung, S. (2022) “PROB: Probabilistic Objectness for Open World Object Detection”, arXiv [Preprint]. Available at: http://arxiv.org/abs/2212.01424v1.
Vancouver
1. Zohar O, Wang K-C, Yeung S (2022) PROB: Probabilistic Objectness for Open World Object Detection. arXiv

BibTeX

@article{zohar2022prob,
  title = {PROB: Probabilistic Objectness for Open World Object Detection},
  author = {Zohar, Orr and Wang, Kuan-Chieh and Yeung, Serena},
  year = {2022},
  journal = {arXiv},
  url = {http://arxiv.org/abs/2212.01424v1},
  eprint = {2212.01424}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: https://creativecommons.org/licenses/by/4.0/