Pith. sign in

REVIEW 4 major objections 4 minor 47 references

Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read This paper argues that in incremental object detection, old-model responses that overlap new-task objects are symbiotic signals carrying reusable old-class cues, and that distilling them at spatial and semantic levels improves retention and

desk verdict A solid, well-ablated IOD method whose central assumption — that high-overlap old-model responses are useful symbiotic signals — is plausible but only indirectly tested; conditional on code and variance. read the letter →

arxiv 2607.13452 v1 pith:QTAM3OMW submitted 2026-07-15 cs.CV cs.AI

classification cs.CVcs.AI
keywords incrementalobjectdetectionknowledgedistillationcatastrophicforgettingsymbiosisquerypartitioningprototyperankingDETRold-newconfusion
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

Incremental object detection trains a detector on new categories while keeping old ones, using only new-task annotations even though old objects still appear in the images. Prior methods discard old-model predictions that overlap new ground-truth boxes, treating them as noise. This paper claims those overlap-heavy responses are actually symbiotic evidence—co-occurrence and occlusion make them carry reusable old-class structure—and that refining and distilling them, rather than dropping them, preserves a unified feature space. On COCO 2017, the proposed method raises AP by 1.9 and 3.0 points over two strong transformer-based baselines on the 70+10 split, with larger gains on old-class retention and reduced old–new confusion.

What carries the argument

The central object is the symbiosis-aware query partition: old-model queries are split into anchors A (high confidence, low IoU with new ground truth), symbiotic queries S (IoU >= 0.7 with new ground truth), and residuals R. The paper's key move is keeping S instead of discarding it. A Consistent Feature Enhancement module refines S with multi-head self-attention and an MLP, using prototype anchors from A to suppress old-class bias, then Spatial Symbiosis Distillation feeds these refined queries through the frozen old decoder and matches them against the new decoder with layer-wise confidence-weighted L2 loss. In parallel, Semantic Symbiosis Distillation builds L2-normalized, confidence-weig

What would settle it

Label the ground-truth boxes that the old model's high-IoU queries overlap and measure the old model's confidence on those boxes. If most of the high-overlap queries correspond to new-class objects with low old-class confidence—meaning the old model is not actually detecting shared old-class structure—then distilling them as old-class knowledge would bias the new model toward old labels. A direct check: in the 70+10 ablation, replace S with a random selection of high-overlap queries and rerun SpSD; if final AP does not drop, the claimed symbiotic content of S is not what drives the gain.

Watch

Extended reading notes

Core claim

The central claim is that the old model's high-overlap predictions—queries whose boxes fit new-class ground truth with IoU at least 0.7—are not discarded evidence but a structured resource. They include new objects misclassified as semantically similar old classes and partially visible old objects still being detected, so they encode spatial and semantic dependencies between old and new categories. The paper shows that distilling these signals through two complementary paths works: Spatial Symbiosis Distillation refines overlap-heavy query features under anchor guidance and aligns them across decoder layers with slot-aligned, confidence-weighted supervision, while Semantic Symbiosis Distilla

Load-bearing premise

The load-bearing premise is that the queries in the symbiotic set—old-model predictions with IoU >= 0.7 against new-task ground truth—carry reusable old-class cues (misclassified new objects and partially visible old objects) rather than being dominated by evidence about the new objects themselves.

Editorial extensions

If this is right

  • Incremental detectors no longer need to treat high-IoU old-model responses as noise; the same responses can be turned into training signal.
  • A unified feature space shared by old and new classes can outperform separation-oriented continual-learning designs in detection.
  • Rank-based alignment of prototype logits preserves old-class semantic structure even when per-instance query correspondences break down.
  • Long-horizon incremental settings and datasets with frequent co-occurrence and occlusion, such as remote-sensing imagery, benefit from this approach.
  • Directly adding high-overlap queries without refinement reduces AP, so the anchor-guided refinement is a necessary component of the method.

Reading between the lines

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

  • Beyond the paper: the same symbiotic-region logic should transfer to class-incremental semantic segmentation, where co-occurring and occluding foreground objects create the same high-overlap gradient conflicts; a slot-aligned distillation over feature maps is a direct testable extension.
  • Beyond the paper: the fixed IoU threshold is a blunt proxy for symbiosis; an adaptive or per-category threshold driven by old-class confidence on new ground truth could separate reusable old-class cues from new-object evidence more cleanly.
  • Beyond the paper: the larger ablation gain from semantic rank alignment than from spatial alignment suggests a pure prototype-rank method without spatial slots might be a stronger baseline than the full pipeline, and would isolate the contribution of each pathway.
  • Beyond the paper: if the symbiotic set truly mixes old-class cues and new-object evidence, a two-branch treatment that distills old-class cues on one branch while suppressing them on the new-class branch could outperform the current single enhancement path.
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 Symbiosis-Inspired Knowledge Distillation (SIKD) for incremental object detection with DETR-style detectors. The method partitions old-model predictions into stable anchors (A), high-overlap symbiotic queries (S), and residual queries (R), then distills anchors with KL/box losses, refines S with a Consistent Feature Enhancement (CFE) module and performs slot-aligned layer-wise distillation, and adds a semantic distillation (SeSD) that aligns soft ranks of confidence-weighted class prototypes. Experiments on COCO 2017 (70+10, 40+40, multi-task) and DIOR report gains over CL-DETR, DyQ-DETR, DCA, and other baselines, with ablations showing each component contributes.

Significance. If the results hold, the paper contributes a new perspective for IOD: instead of discarding old-model predictions with high overlap to new-class ground truth, SIKD treats them as useful 'symbiotic' signals and shows that, with appropriate refinement, they can improve retention and reduce old–new confusion. The empirical study is reasonably thorough, covering multiple benchmarks, multi-step settings, ablations, and hyperparameter sensitivity, and the code appears reproducible (though not provided). The main risk is that the central assumption about the semantic value of the S set is only indirectly validated, and several comparisons lack repeated-seed statistics, so the magnitude of the claimed gains—especially on 40+40—needs stronger support.

major comments (4)
  1. [§4.3, Table 5] The α ablation rows appear to be inconsistent with the full model. The α=1 row reports All AP 42.3, which exactly matches the SpSD-only result in Table 3 (Idx 3), while the full SIKD (α=1, β=6) is 44.3 in Table 3 (Idx 5). If the α sweep was performed without SeSD, then the text's conclusion that 'α=1 achieves the highest All AP' does not support the choice of α for the full model. Please clarify the configuration used for the α rows and provide an α sweep on the full model.
  2. [§3.3, Eq. (3); Fig. 2] The definition of S (all queries with IoU ≥ τ against new-class ground truth) is justified by the claim that these queries are 'symbiotic' old-class cues. However, Fig. 2 only shows confidence distributions and per-class misclassification proportions; it does not directly establish semantic relatedness between the predicted old class and the true new class. The ablation in Table 3 (Idx 2 vs 3) shows that naive distillation of S hurts, implying S is not inherently benign—the benefit appears only after CFE transforms it. To verify the load-bearing mechanism, please (i) report the distribution of predicted old classes for S queries and their semantic similarity (e.g., WordNet or embedding distance) to the corresponding new-class ground truth, (ii) include an ablation where CFE is applied to a matched set of high-IoU queries not defined by S, and (iii) report what fraction of S queries corre
  3. [§4.1, §4.2, Tables 1–2] All comparisons are based on single runs, and no error bars or repeated-seed results are reported. On the 40+40 split the gains over DyQ-DETR and DCA are only 0.9 and 0.5 AP, which is within typical run-to-run variation for DETR-style detectors. In addition, hyperparameters (α, β, γ, τ) are fixed across settings but selected on COCO 70+10. Please report mean ± std over at least 3 seeds for the main settings, and clarify whether the hyperparameters were tuned on a validation split or the test set, or show cross-setting sensitivity.
  4. [§4.2, Table 2] In the multi-task settings, SIKD's base-phase AP is 45.4, whereas DCA's is 48.0. The final-phase AP of SIKD is 38.1 vs DCA's 37.2, so the improvement could partly reflect a lower starting point rather than superior incremental learning. The text claims 'superior retention of prior knowledge,' but the table reports only overall AP. Please also report old-class AP (or a forgetting metric) and new-class AP per phase so that the retention–plasticity balance is directly evidenced.
minor comments (4)
  1. [§3.5, Eq. (4)] The update in Eq. (4) applies to all queries not in A, including residual queries R. The text explains that R carries 'weak relational cues,' but it is unclear whether the CFE loss in Eq. (6) also covers R or only S. Please clarify the role of R in the enhancement and whether any loss is applied to it.
  2. [§4.3] The text refers to 'Fig. 4(b) and (c)', but Figure 4 only contains panels (a) and (b). Please correct the reference or add the missing panel.
  3. [§3.5, Eq. (7)] If the set A is empty in a mini-batch (which can happen for the 19+1 setting), the anchor loss L_A in Eq. (7) is undefined. Similarly, L_CFE in Eq. (6) is undefined if S is empty. Please specify a fallback (e.g., skip the corresponding loss for that batch).
  4. [§4.1] The text says 'we use fixed hyperparameters across all settings,' but Table 5 shows hyperparameter selection on COCO 70+10. This wording is contradictory; please rephrase to indicate that hyperparameters are tuned once on the primary benchmark and then kept fixed.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: SIKD is an empirically validated distillation method whose headline results are external benchmark comparisons; the 'symbiotic set' is defined by geometric overlap and validated by ablation, not by definition.

full rationale

This is an empirical methods paper, not a derivation, so equation-level circularity is not at issue. The symbiotic set S (Eq. 3) is defined purely by IoU >= tau against new-task ground truth; the claim that these queries carry reusable old-class cues is an empirical hypothesis supported by the study in Fig. 2 and by end-task accuracy, not an identity. The distillation objectives (Eqs. 6/7/9/13) align the new model to the frozen old model — the standard mechanism of knowledge distillation in IOD — and the gains are measured against external baselines (DyQ-DETR, DCA, CL-DETR) on held-out COCO/DIOR validation. Crucially, the ablation (Table 3, Idx 2 vs 3) shows that naively distilling the high-overlap responses decreases AP by 0.1, so the reported benefit of SpSD is not forced by construction; CFE (Eqs. 4-6) is what makes it useful. Hyperparameters gamma, tau, alpha, beta are tuned and honestly reported (Tables 5 and 7) — this is standard model selection, not a fitted input renamed as a prediction. Self-citations (Cheng et al. 2024/2026, Xu et al. 2025/2026, He et al. 2025a/b/2026, Wang et al. 2025a/b) appear only as related-work context and future-work suggestions, and no load-bearing claim is justified solely by these citations. The concern that S queries may not be semantically related to the predicted old class is an untested-assumption (robustness) risk, not circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No ontologically new physical entities are posited; CFE is an architectural module whose evidence is internal ablations. The main burden, therefore, sits on the partitioning assumption about high-overlap queries and on the standard DETR/MMDetection fidelity assumptions.

free parameters (5)
  • alpha (α) = 1.0
    Balance weight for layer-wise spatial distillation; chosen from ablation on COCO 70+10 (Table 5) for best All AP.
  • beta (β) = 6.0
    Balance weight for SeSD; chosen from ablation on COCO 70+10 (Table 5) for best Avg AP / favorable old-new trade-off.
  • gamma (γ) = 0.4
    Old-class confidence threshold in query partitioning; sensitivity analysis in Appendix C (Table 7) selects 0.4.
  • tau (τ) = 0.7
    IoU threshold defining symbiotic/overlap-set queries; sensitivity analysis in Appendix C (Table 7) selects 0.7.
  • lambda_1, lambda_2 = 5.0, 2.0
    Box distillation weights for anchors, inherited from KD-DETR (Wang et al. 2024) rather than tuned here; still hyperparameters affecting the loss.
assumptions (4)
  • ad hoc to paper Old-model high-overlap responses encode reusable old-class cues rather than pure noise.
    Introduced in Sec. 3.3 and used to define the symbiotic set S; no independent proof is given.
  • domain assumption Soft-rank alignment of old-class logits preserves semantic topology and reduces forgetting.
    Invoked in Sec. 3.6 and Appendix B, following Tao et al. 2020 and Liu et al. 2022.
  • domain assumption Deformable DETR query slots are alignable across the old and new decoder layers.
    Used for SpSD's slot-aligned, layer-wise distillation in Sec. 3.5 (Eq. 8-9).
  • domain assumption The pseudo-labeling threshold protocol (γ=0.4, τ=0.7) yields a sufficiently reliable anchor set A.
    Standard IOD practice, used in Sec. 3.3 and Sec. 4.1, not independently justified beyond sensitivity analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection." pith.science (2026). https://pith.science/paper/QTAM3OMW

@misc{pith2026260713452,
  author       = {Pith},
  title        = {Pith review of: Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QTAM3OMW}},
  note         = {Machine review of arXiv:2607.13452}
}
read the original abstract

Incremental object detection (IOD) aims to extend detectors to new categories while retaining previously acquired knowledge. Existing methods often adopt a class incremental learning perspective, separating feature spaces to sharpen decision boundaries. However, this separation-oriented paradigm may overlook object symbiosis in detection, where co-occurrence and occlusion introduce spatial and semantic dependencies that benefit from shared representations. Ignoring these dependencies distorts the shared representations, exacerbates confusion between old and new classes, and accelerates catastrophic forgetting. To address this, we propose Symbiosis-Inspired Knowledge Distillation (SIKD), which explicitly leverages object symbiosis at two complementary levels. Spatial Symbiosis Distillation (SpSD) focuses on symbiotic regions where the old model responds with high overlap to objects in the new task. It preserves generalizable old class cues, suppresses class-specific bias and redundancy, and distills the refined evidence to the new model at matched spatial locations with slot-aligned supervision. Semantic Symbiosis Distillation (SeSD) maintains class level structure by forming confidence weighted prototypes for old classes and aligning their inter class soft ranks over the old class logits, which stabilizes the semantic topology during adaptation. Extensive experiments demonstrate the effectiveness and superiority of the proposed method.

Figures

Figures reproduced from arXiv: 2607.13452 by the authors.

Figure 1
Figure 1. Illustration of (a) object symbiosis in IOD, (b) existing methods in new task, and (c) our method in new task. In (b) and (c), arrows indicate how regional features are classified. In (c), the new-task class apple shares coarse features with the old class orange while also retaining class-specific cues. stage detectors (Ge et al., 2021; Tian et al., 2019) and end-to-end transformer architectures (Zhu et al., 2020; L… view at source ↗
Figure 2
Figure 2. Statistical analysis on COCO 2017 under the 70+10 setting, using old-detector predictions with IoU > 0.7 to new-class ground truth. (a) Confidence distribution and mean IoU of the old detector’s old-class predictions on new-class ground-truth instances. (b) Per-class proportion of new-class ground-truth instances misclassified as old classes by the old detector. techniques (Wang et al., 2025c). DCA (Zhang et al., 20… view at source ↗
Figure 3
Figure 3. Overview of our proposed SIKD. (a) Training pipeline. The frozen old model Mt−1 produces queries Q t−1 on D t . CFE refines queries of symbiotic regions under anchor-prototype guidance, yielding Q E that reduces old-class bias and removes redundancy. SpSD distills anchor logits and boxes, which enforces confidence-weighted, layer-wise logit consistency over all queries. SeSD builds confidence-weighted, L2-normalized… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Comparison of confusion matrices for old and new Classes between the baseline and our method on COCO (70+10). AP by 0.4 and Old AP by 0.5, while reducing New AP by 0.6. Setting β = 6 yields the best task-averaged AP (42.0) and provides a favorable trade-off between ret…
Figure 6
Figure 6. Figure 6: t-SNE visualization of object features on the COCO 2017 validation set. (a) Baseline after the final incremental task on COCO 2017 (70+10) setting, (b) Joint Training as an upper-bound model trained once on the union of old and new classes with full annotations (non-in…
Figure 7
Figure 7. Figure 7: Visualization of the CFE module. For each detection, class id and confidence are shown as “old→new” with the left value before CFE and the right value after CFE. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Qualitative visualization of detection predictions from the baseline and our method (SIKD) on COCO 2017 (70+10) setting. In (a), oven is an old class and microwave is a new class. In (b), carrot is an old class and apple is a new class. In (c), both orange and person a…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 5 linked inside Pith

  1. [1]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Learning without forgetting , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2017 , publisher=

  2. [2]

    Proceedings of the 4th ACM/IEEE Symposium on Edge Computing , pages=

    RILOD: Near real-time incremental learning for object detection at the edge , author=. Proceedings of the 4th ACM/IEEE Symposium on Edge Computing , pages=

  3. [3]

    Computer vision and image understanding , volume=

    Sid: Incremental learning for anchor-free object detection via selective and inter-related distillation , author=. Computer vision and image understanding , volume=. 2021 , publisher=

  4. [4]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Continual detection transformer for incremental object detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  5. [5]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Overcoming catastrophic forgetting in incremental object detection via elastic response distillation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  6. [6]

    ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Dynamic Object Queries for Transformer-based Incremental Object Detection , author=. ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2025 , organization=

  7. [7]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Learning task-aware language-image representation for class-incremental object detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Kd-detr: Knowledge distillation for detection transformer with consistent distillation points sampling , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Show all 47 references
  1. [9]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    DCA: Dividing and Conquering Amnesia in Incremental Object Detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  2. [10]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Sddgr: Stable diffusion-based deep generative replay for class incremental object detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  3. [11]

    Computer vision--ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13 , pages=

    Microsoft coco: Common objects in context , author=. Computer vision--ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13 , pages=. 2014 , organization=

  4. [12]

    arXiv preprint arXiv:1906.07155 , year=

    MMDetection: Open mmlab detection toolbox and benchmark , author=. arXiv preprint arXiv:1906.07155 , year=

  5. [13]

    Proceedings of the national academy of sciences , volume=

    Overcoming catastrophic forgetting in neural networks , author=. Proceedings of the national academy of sciences , volume=. 2017 , publisher=

  6. [14]

    arXiv preprint arXiv:2506.12351 , year=

    EKPC: Elastic Knowledge Preservation and Compensation for Class-Incremental Learning , author=. arXiv preprint arXiv:2506.12351 , year=

  7. [15]

    arXiv preprint arXiv:2505.13997 , year=

    StPR: Spatiotemporal Preservation and Routing for Exemplar-Free Video Class-Incremental Learning , author=. arXiv preprint arXiv:2505.13997 , year=

  8. [16]

    arXiv preprint arXiv:2508.01579 , year=

    Harnessing Textual Semantic Priors for Knowledge Transfer and Refinement in CLIP-Driven Continual Learning , author=. arXiv preprint arXiv:2508.01579 , year=

  9. [17]

    arXiv preprint arXiv:2507.09471 , year=

    CKAA: Cross-subspace Knowledge Alignment and Aggregation for Robust Continual Learning , author=. arXiv preprint arXiv:2507.09471 , year=

  10. [18]

    Advances in neural information processing systems , volume=

    Online continual learning with maximal interfered retrieval , author=. Advances in neural information processing systems , volume=

  11. [19]

    Advances in neural information processing systems , volume=

    Gradient based sample selection for online continual learning , author=. Advances in neural information processing systems , volume=

  12. [20]

    IEEE Transactions on Image Processing , year=

    Balanced destruction-reconstruction dynamics for memory-replay class incremental learning , author=. IEEE Transactions on Image Processing , year=

  13. [21]

    Advances in neural information processing systems , volume=

    Continual learning with node-importance based adaptive group sparse regularization , author=. Advances in neural information processing systems , volume=

  14. [22]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Der: Dynamically expandable representation for class incremental learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  15. [23]

    International conference on machine learning , pages=

    Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting , author=. International conference on machine learning , pages=. 2019 , organization=

  16. [24]

    Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , pages=

    icarl: Incremental classifier and representation learning , author=. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , pages=

  17. [25]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  18. [26]

    The Thirteenth International Conference on Learning Representations , year=

    PseDet: Revisiting the Power of Pseudo Label in Incremental Object Detection , author=. The Thirteenth International Conference on Learning Representations , year=

  19. [27]

    European Conference on Computer Vision , pages=

    Bridge past and future: Overcoming information asymmetry in incremental object detection , author=. European Conference on Computer Vision , pages=. 2024 , organization=

  20. [28]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Faster R-CNN: Towards real-time object detection with region proposal networks , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2016 , publisher=

  21. [29]

    arXiv preprint arXiv:2107.08430 , year=

    Yolox: Exceeding yolo series in 2021 , author=. arXiv preprint arXiv:2107.08430 , year=

  22. [30]

    Proceedings of the IEEE international conference on computer vision , pages=

    Fast r-cnn , author=. Proceedings of the IEEE international conference on computer vision , pages=

  23. [31]

    arXiv preprint arXiv:2010.04159 , year=

    Deformable detr: Deformable transformers for end-to-end object detection , author=. arXiv preprint arXiv:2010.04159 , year=

  24. [32]

    European conference on computer vision , pages=

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection , author=. European conference on computer vision , pages=. 2024 , organization=

  25. [33]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Alleviating catastrophic forgetting of incremental object detection via within-class and between-class knowledge distillation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  26. [34]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Class-incremental learning: A survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  27. [35]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Up-detr: Unsupervised pre-training for object detection with transformers , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  28. [36]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Fcos: Fully convolutional one-stage object detection , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  29. [37]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Class-incremental learning: survey and performance evaluation on image classification , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2022 , publisher=

  30. [38]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Fcs: Feature calibration and separation for non-exemplar class incremental learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  31. [39]

    European conference on computer vision , pages=

    Topology-preserving class-incremental learning , author=. European conference on computer vision , pages=. 2020 , organization=

  32. [40]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    Model behavior preserving for class-incremental learning , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2022 , publisher=

  33. [41]

    ISPRS journal of photogrammetry and remote sensing , volume=

    Object detection in optical remote sensing images: A survey and a new benchmark , author=. ISPRS journal of photogrammetry and remote sensing , volume=. 2020 , publisher=

  34. [42]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Augmented box replay: Overcoming foreground shift for incremental object detection , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  35. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Disentangled prompt representation for domain generalization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  36. [44]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Adversarial domain prompt tuning and generation for single domain generalization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  37. [45]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Prompt Disentanglement via Language Guidance and Representation Alignment for Domain Generalization , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  38. [46]

    Reasoning-Driven Multimodal

    Xu, Zhipeng and Wang, Zilong and Jiang, Xinyang and Li, Dongsheng and Cheng, De and Wang, Nannan , booktitle =. Reasoning-Driven Multimodal. 2026 , url =

  39. [47]

    Forty-third International Conference on Machine Learning , year=

    Task-Driven Subspace Decomposition for Knowledge Sharing and Isolation in LoRA-based Continual Learning , author=. Forty-third International Conference on Machine Learning , year=

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.