Pith. sign in

REVIEW 2 major objections 6 minor 26 references

DEUS separates known and unknown object proposals with two energy subspaces and protects old-versus-new class learning during memory replay, raising unknown recall while keeping known-class accuracy competitive.

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

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

T0 review · grok-4.5

2026-07-13 15:26 UTC pith:FET7NMN6

load-bearing objection Solid empirical OWOD paper: dual fixed-ETF energy modules give large unknown-recall gains; main residual risk is the usual pseudo-label dependence. the 2 major comments →

arxiv 2603.29954 v2 pith:FET7NMN6 submitted 2026-03-31 cs.CV

Detecting Unknown Objects via Energy-based Separation for Open World Object Detection

Jun-Woo Heo , Keonhee Park , Gyeong-Moon Park This is my paper
classification cs.CV
keywords open-world object detectionenergy-based modelsEquiangular Tight Frameunknown object detectioncatastrophic forgettingmemory replaycontinual learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Open-world object detection asks a model to keep learning new labeled classes without forgetting old ones, while also spotting unlabeled “unknown” objects it has never seen. Earlier methods mostly decide what is unknown by looking at how strongly the known-class head fires, so unknown objects never get their own clean representation and often collapse into background. Memory replay helps keep old classes alive but lets old and new classifiers interfere with each other as the label set grows. DEUS attacks both problems at once. It builds two fixed, geometrically orthogonal subspaces from a Simplex Equiangular Tight Frame and trains proposals so that known objects score high only in the known subspace, unknown proposals score high only in the unknown subspace, and background sits near the boundary. A second energy loss then splits the classifier into previous-task and current-task heads and forces each sample to prefer its own head, cutting cross-talk during replay. On standard OWOD benchmarks the result is a large jump in unknown recall together with stable or better known-class mAP as more tasks arrive.

Core claim

Modeling separate known and unknown Simplex-ETF subspaces and scoring proposals by the energy difference between those subspaces lets a detector learn distinct unknown representations; pairing that with an energy-based preference between previous-task and current-task classifier heads removes the usual trade-off between old-class retention and new-class acquisition.

What carries the argument

ETF-Subspace Unknown Separation (EUS): fixed known and unknown Simplex-ETF bases produce two Helmholtz free energies; the signed offset between those energies, together with a margin loss and a three-way focal loss, drives known, unknown, and background proposals into distinct geometric regions. Energy-based Known Distinction (EKD) then applies the same energy idea to previous-versus-current classifier heads during replay.

Load-bearing premise

That two fixed, non-learnable ETF subspaces plus a simple energy-margin loss can cleanly separate known, unknown, and background even when the only labels for unknowns are the detector’s own pseudo-labels.

What would settle it

Train the identical detector with and without the two ETF subspaces on the same OWOD task sequence; if the large gain in unknown recall disappears once the subspaces are removed (or replaced by a single known-only energy), the central geometric claim fails.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper proposes DEUS for Open World Object Detection, combining ETF-Subspace Unknown Separation (EUS) and an Energy-based Known Distinction (EKD) loss. EUS constructs fixed Simplex-ETF known and unknown subspaces (Eqs. 4–6), defines dual Helmholtz free energies (Eqs. 7–8), and trains with an energy-margin loss plus a subspace focal loss (Eqs. 11–13), with inference calibration of the unknown logit via the standardized unknown offset (Eq. 14). EKD splits the known classifier into previous/current sub-classifiers during memory replay and applies a pairwise energy preference loss (Eqs. 15–16) to reduce cross-influence. Built on OrthogonalDet, DEUS is evaluated on M-OWODB, S-OWODB, and a new RS-OWODB split, reporting large gains in unknown recall (e.g., U-Rec 65.1/66.2/69.0 on M-OWODB Tasks 1–3) with competitive known mAP and best H-Scores, supported by ablations (Table 3), subspace-score and energy heatmaps (Fig. 3), PCA (Fig. 1a), and qualitative examples (Fig. 4).

Significance. If the reported gains hold under full reproducibility, DEUS is a meaningful advance for OWOD: it is the first dual-subspace energy formulation that explicitly models unknown geometry rather than relying only on known-class energies, and EKD gives a simple, energy-based regularizer for the old/new trade-off under replay. The empirical package is strong—standard M-/S-OWODB protocols, a new remote-sensing split (RS-OWODB), component ablations, geometry visualizations, and modest overhead numbers. The U-Rec improvements are large enough to matter for open-world detectors while known mAP remains competitive as the class set grows. Strengths that should be credited include the clear ablation isolation of EUS vs. EKD (Table 3), the per-class subspace-score analysis (Fig. 3a), and the explicit RS-OWODB generalization check.

major comments (2)
  1. [Sec. 4.1, Table 3, Appendix A] Sec. 4.1 and Table 3 vs. Table 1: the ablation baseline already incorporates an “improved” pseudo-labeling process (dynamic scaling by known-class count and noise filtering) that is only sketched and deferred to Appendix A. Because this process changes the training targets for unknowns and is shared by the DEUS rows, the main text should (i) fully specify the selection/filtering rules and (ii) report an OrthogonalDet-style baseline both with and without the improved pseudo-labeler, so that the U-Rec jump attributable to EUS is cleanly separated from the pseudo-label schedule. Without that isolation, the central unknown-detection claim is harder to audit.
  2. [Sec. 3.3, Eqs. 4–6, 11] Sec. 3.3, Eqs. (4)–(6) and (11): EUS rests on fixed, non-learnable Simplex-ETF bases with K=128 (64/64 split) and a margin m in L_energy. No sensitivity to K or m is reported, nor any comparison to learnable or randomly fixed bases. Given that unknown supervision is only the detector’s own pseudo-labels, a short sensitivity study (and, if possible, a random-orthonormal control) is needed to show that the claimed known/unknown/background geometry is robust rather than specific to one hyperparameter setting.
minor comments (6)
  1. [Figure 1a] Fig. 1a PCA axes and color legend are hard to read in grayscale; please enlarge markers and state the number of proposals sampled per category.
  2. [Sec. 3.3, Eq. 14] Eq. (14): the standardization of Δu is per-image; clarify whether μ and σ are computed over all proposals or only high-objectness proposals, and whether this is done at train time as well as inference.
  3. [Table 1] Table 1: several † footnotes note corrected M-OWODB annotations; state explicitly which prior numbers were re-run under the same corrected labels so that ranking comparisons remain fair.
  4. [Sec. 3.2–3.3] Notation: z_cls is said to have C+1 nodes including an unknown node, yet EUS also injects a calibrated unknown logit z′_u; a short paragraph on how these two unknown pathways interact at training vs. inference would help.
  5. [Sec. 2.2] Related Work (Sec. 2.2) could briefly contrast dual-subspace energy with single-space energy OOD methods beyond the OWOD citations already listed.
  6. [Abstract / Sec. 1] Minor typos: “DetectingUnknowns” spacing in the abstract acronym expansion; “i.e., catastrophic for-getting” line break; ensure consistent hyphenation of “open-world” vs. “Open World”.

Circularity Check

0 steps flagged

No significant circularity: purely empirical OWOD method whose losses and ETF construction are taken from external literature and validated on external benchmarks.

full rationale

DEUS defines two new losses (EUS via fixed non-learnable Simplex-ETF subspaces plus energy-margin/subspace terms, Eqs. 4–13; EKD via partitioned classifier energies, Eqs. 15–16) and reports empirical mAP/U-Rec/H-Score gains on M-OWODB, S-OWODB and RS-OWODB. There is no first-principles derivation, no fitted parameter that is later re-presented as a prediction, no uniqueness theorem, and no algebraic identity that reduces a claimed result to its own inputs. The ETF construction and Helmholtz free-energy scores are standard citations (Papyan et al., Liu et al.); the single overlapping-author citation (Park et al. 2025) appears only in related work and is not load-bearing. Ablations (Table 3), subspace-score diagnostics (Fig. 3a) and qualitative examples supply independent empirical support. The paper is therefore self-contained against external benchmarks and exhibits zero circularity of the kinds enumerated.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central performance claim rests on a small set of geometric and training assumptions plus a handful of free hyper-parameters; no new physical entities are postulated. The ETF bases are taken as fixed non-learnable objects whose equiangular property is assumed to induce useful separation once the energy losses are applied.

free parameters (4)
  • K (number of ETF basis vectors) = 128
    Set to 128 (64 known + 64 unknown) without ablation in the main text; controls the dimensionality of the two energy modules.
  • margin m in L_energy
    Hyper-parameter that enforces the minimum gap between known and unknown subspace scores; value not reported in the main body.
  • loss weights for L_EUS and L_EKD = 1.0
    Both fixed to 1.0; no sensitivity study shown.
  • number of pseudo-unknown labels per image
    Dynamically scaled by known-class count and filtered; the exact schedule is only described as “improved” in Appendix A.
axioms (4)
  • ad hoc to paper A fixed Simplex ETF split into two non-overlapping subspaces yields geometrically clean known and unknown energy modules that can be used without further learning of the bases.
    Invoked in Sec. 3.3, Eqs. 4–6; the paper treats the bases as non-learnable and relies on the subsequent energy losses to align features.
  • domain assumption Pseudo-labels generated by the detector itself (with the improved filtering) are sufficiently accurate targets for the unknown subspace.
    Standard OWOD practice, but load-bearing for EUS training; described in Sec. 3.2 and Appendix A.
  • domain assumption Helmholtz free energy (negative log-sum-exp of projections) is a valid affinity score for both the ETF subspaces and the split classifiers.
    Taken from Liu et al. energy-based OOD work and applied in Eqs. 7–9 and 15.
  • domain assumption Memory-replay exemplars plus the EKD pairwise energy loss are sufficient to control cross-influence between previous and current classes.
    Sec. 3.4; the loss is applied only during replay phases.
invented entities (2)
  • ETF-Subspace Unknown Separation (EUS) dual energy modules no independent evidence
    purpose: Provide separate known and unknown geometric spaces so that energy scores can pull proposals into the correct region and push background to the boundary.
    Constructed in Sec. 3.3 from a fixed Simplex ETF; no independent existence outside the paper’s training objective.
  • Energy-based Known Distinction (EKD) loss no independent evidence
    purpose: Force previous-task proposals to score higher on the previous sub-classifier than on the current one (and vice versa) during memory replay.
    Defined in Sec. 3.4, Eq. 16; purely a training regularizer introduced by the paper.

pith-pipeline@v1.1.0-grok45 · 17359 in / 3076 out tokens · 28567 ms · 2026-07-13T15:26:49.544561+00:00 · methodology

0 comments
read the original abstract

In this work, we tackle the problem of Open World Object Detection (OWOD). This challenging scenario requires the detector to incrementally learn to classify known objects without forgetting while identifying unknown objects without supervision. Previous OWOD methods have enhanced the unknown discovery process and employed memory replay to mitigate catastrophic forgetting. However, since existing methods heavily rely on the detector's known class predictions for detecting unknown objects, they struggle to effectively learn and recognize unknown object representations. Moreover, while memory replay mitigates forgetting of old classes, it often sacrifices the knowledge of newly learned classes. To resolve these limitations, we propose DEUS (Detecting Unknowns via energy-based Separation), a novel framework that addresses the challenges of Open World Object Detection. DEUS consists of Equiangular Tight Frame (ETF)-Subspace Unknown Separation (EUS) and an Energy-based Known Distinction (EKD) loss. EUS leverages ETF-based geometric properties to create orthogonal subspaces, enabling cleaner separation between known and unknown object representations. Unlike prior energy-based approaches that consider only the known space, EUS utilizes energies from both spaces to better capture distinct patterns of unknown objects. Furthermore, EKD loss enforces the separation between previous and current classifiers, thus minimizing knowledge interference between previous and newly learned classes during memory replay. We thoroughly validate DEUS on OWOD benchmarks, demonstrating outstanding performance improvements in unknown detection while maintaining competitive known class performance.

discussion (0)

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

Reference graph

Works this paper leans on

26 extracted references · 1 linked inside Pith

  1. [1]

    On the effectiveness of lipschitz-driven rehearsal in continual learning.Advances in Neural Information Processing Systems, 35:31886–31901, 2022

    Lorenzo Bonicelli, Matteo Boschini, Angelo Porrello, Con- cetto Spampinato, and Simone Calderara. On the effectiveness of lipschitz-driven rehearsal in continual learning.Advances in Neural Information Processing Systems, 35:31886–31901, 2022

    2022

  2. [2]

    MMDetection: Open mmlab detection toolbox and benchmark.arXiv preprint arXiv:1906.07155, 2019

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tian- heng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and...

    Pith/arXiv arXiv 1906

  3. [3]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13678–13688, 2022

    2022

  4. [4]

    You only look at one sequence: Rethinking transformer in vision through object detection.Advances in Neural Information Processing Systems, 34:26183–26197, 2021

    Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object detection.Advances in Neural Information Processing Systems, 34:26183–26197, 2021

    2021

  5. [5]

    Fast r-cnn

    Ross Girshick. Fast r-cnn. InInternational Conference on Computer Vision (ICCV), 2015

    2015

  6. [6]

    Rich feature hierarchies for accurate object detection and semantic segmentation

    Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014

    2014

  7. [7]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9235–9244, 2022

    2022

  8. [8]

    Towards open world object detec- tion

    KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vi- neeth N Balasubramanian. Towards open world object detec- tion. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 5830–5840, 2021

    2021

  9. [9]

    Object detection in optical remote sensing images: A survey and a new benchmark, 2019

    Ke Li, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark, 2019

    2019

  10. [10]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3230–3239, 2023

    2023

  11. [11]

    Energy-based out-of-distribution detection.Advances in neu- ral information processing systems, 33:21464–21475, 2020

    Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection.Advances in neu- ral information processing systems, 33:21464–21475, 2020

    2020

  12. [12]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19681– 19690, 2023

    2023

  13. [13]

    Annealing- based label-transfer learning for open world object detection

    Yuqing Ma, Hainan Li, Zhange Zhang, Jinyang Guo, Shang- hang Zhang, Ruihao Gong, and Xianglong Liu. Annealing- based label-transfer learning for open world object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11454–11463, 2023

    2023

  14. [14]

    The stability-plasticity dilemma: Investigating the continuum from catastrophic forgetting to age-limited learning effects, 2013

    Martial Mermillod, Aur ´elia Bugaiska, and Patrick Bonin. The stability-plasticity dilemma: Investigating the continuum from catastrophic forgetting to age-limited learning effects, 2013

    2013

  15. [15]

    An end-to- end transformer model for 3d object detection

    Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to- end transformer model for 3d object detection. InProceedings of the IEEE/CVF international conference on computer vision, pages 2906–2917, 2021

    2021

  16. [16]

    Prevalence of neural collapse during the terminal phase of deep learning training.Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020

    Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training.Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020

    2020

  17. [17]

    Online continuous generalized category discov- ery

    Keon-Hee Park, Hakyung Lee, Kyungwoo Song, and Gyeong- Moon Park. Online continuous generalized category discov- ery. InEuropean Conference on Computer Vision, pages 53–69. Springer, 2025

    2025

  18. [18]

    Generalized in- tersection over union: A metric and a loss for bounding box regression

    Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized in- tersection over union: A metric and a loss for bounding box regression. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658–666, 2019

    2019

  19. [19]

    Focal loss for dense object detection

    T-YLPG Ross and GKHP Doll´ar. Focal loss for dense object detection. Inproceedings of the IEEE conference on computer vision and pattern recognition, pages 2980–2988, 2017

    2017

  20. [20]

    Sparse r-cnn: End-to-end object detection with learnable proposals

    Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. InProceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 14454–14463, 2021

    2021

  21. [21]

    Exploring or- thogonality in open world object detection

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

    2024

  22. [22]

    Re- hearsal revealed: The limits and merits of revisiting samples in continual learning

    Eli Verwimp, Matthias De Lange, and Tinne Tuytelaars. Re- hearsal revealed: The limits and merits of revisiting samples in continual learning. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 9385–9394, 2021

    2021

  23. [23]

    Random boxes are open-world object detectors

    Yanghao Wang, Zhongqi Yue, Xian-Sheng Hua, and Hanwang Zhang. Random boxes are open-world object detectors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6233–6243, 2023

    2023

  24. [24]

    O1o: Grouping of known classes to identify unknown objects as odd-one-out

    Mısra Yavuz and Fatma G¨uney. O1o: Grouping of known classes to identify unknown objects as odd-one-out. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 614–629, 2024

    2024

  25. [25]

    Shan Zhang, Yao Ni, Jinhao Du, Yuan Xue, Philip H. S. Torr, Piotr Koniusz, and Anton van den Hengel. Open-world object- ness modeling unifies novel object detection.2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 30332–30342, 2025

    2025

  26. [26]

    Prob: Prob- abilistic objectness for open world object detection

    Orr Zohar, Kuan-Chieh Wang, and Serena Yeung. Prob: Prob- abilistic objectness for open world object detection. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11444–11453, 2023

    2023