Pith. sign in

REVIEW 3 major objections 4 minor 41 references

iFAN: Inference-Aware Learning for Plain Mask Transformers

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read iFAN shows that mask transformers can be trained to choose better query masks at inference, adding 1.20 PQ on average without extra inference cost.

desk verdict Useful training recipe for mask transformers; the gains survive, but the abstract's AP average is contradicted by the paper's own corrected tables. read the letter →

arxiv 2608.03216 v1 pith:JSCDJQFM submitted 2026-08-04 cs.CV

Fang Li , Yu He , Haoyang Tong , Lichen Ma , Jingling Fu , Wenxiao Fan , Tongxuan Liu , Luohang Liu
show 2 more authors
Ke Zhang Junshi Huang
This is my paper
classification cs.CV
keywords query-basedsegmentationmasktransformerspanopticinstancesemanticself-distillationqualityrankinginference-awarelearning
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

Query-based mask transformers decide their final segmentation by letting every query compete pixel-by-pixel at the last layer, but that competition is never explicitly optimized during training. This paper argues that this creates two correctable mismatches: the query with the highest probability–mask score is often not the one with the most accurate mask, and the final layer frequently discards better predictions made by intermediate layers. The proposed framework, iFAN, adds two training-only objectives—Adjusted Probability-Mask Ranking (APMR) and Cross-Layer Self-Distillation (CLSD)—that calibrate the scores used in the competition and transfer strong intermediate predictions into the final layer. Because both objectives disappear at inference, the method claims consistent accuracy gains across panoptic, instance, and semantic segmentation—averaging 1.20 PQ, 1.30 AP, and 0.63 mIoU—with negligible added parameters, FLOPs, or latency. If true, iFAN is a drop-in training upgrade for plain mask transformers.

What carries the argument

The central object is the adjusted probability-mask score map $\hat{S}_i = (\hat{q}_i \hat{p}_i(c_k))^\gamma \hat{M}_i$, where $\hat{q}_i$ is a predicted global mask-quality score learned by regression to soft-IoU. It carries the argument because it replaces the raw probability-mask product as the quantity that determines query competition; APMR then ranks matched queries above hard negatives on this map, and CLSD distills this same map from better intermediate layers into the final layer. The ranking and distillation losses are the mechanism; the shared quality head is the only new parameter-bearing component.

What would settle it

Measure the actual wall-clock latency, parameter count, and GFLOPs of an iFAN-trained model (including the quality head and the adjusted score computation) against its matched baseline; if the iFAN model is visibly slower or larger in an apples-to-apples benchmark, the central efficiency claim fails.

Watch

Extended reading notes

Core claim

iFAN’s central claim is that the inference-time query competition of a mask transformer can be made consistent with true mask quality without touching the inference pipeline. APMR attaches a lightweight shared quality head to every query, supervised by the soft-IoU between the predicted mask and its matched ground-truth target, and uses the adjusted score map $\hat{S}_i = (\hat{q}_i \hat{p}_i(c_k))^\gamma \hat{M}_i$ with $\gamma=2$ as the competition score. A pairwise ranking loss then forces the matched query’s adjusted score above the adjusted scores of hard-negative unmatched queries—queries that confidently predict the target class but overlap the target poorly—with gradients stopped fro

Load-bearing premise

The claim that iFAN adds negligible inference cost assumes the shared-configuration accounting is valid: the paper lists the baseline’s parameters, GFLOPs, and FPS in iFAN rows rather than measuring the model with the added quality head and adjusted scoring, so an independent timing could contradict “negligible.”

Editorial extensions

If this is right

  • Plain mask transformers trained with iFAN improve on all three segmentation tasks against their matched baselines: average +1.20 PQ, +1.30 AP, and +0.63 mIoU across COCO, ADE20K, and Cityscapes.
  • The gains hold across two architectures, backbones from ViT-S to ViT-G, and input resolutions, while parameters, GFLOPs, and FPS are reported as unchanged under shared-configuration accounting.
  • APMR reduces the fraction of inference selections where the highest-scoring query disagrees with the best-quality query from 64.45% to 27.49%, and the high-confidence/low-IoU rate from 29.27% to 11.50%.
  • CLSD shifts the number of classes whose best prediction is produced at the final layer from 53 to 98 out of 133, meaning intermediate-layer knowledge is preserved without multi-layer inference.
  • Smaller backbones gain more from iFAN (1.6 PQ at ViT-S vs 0.8 PQ at ViT-G), suggesting the correction matters most when the model has fewer parameters.

Reading between the lines

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

  • The framework’s logic is not tied to plain mask transformers; APMR and CLSD could be bolted onto any query-based segmenter or detector whose final output is chosen by a per-query score, provided a quality signal such as soft-IoU is available during training.
  • A direct test of the quality head’s generality would be to measure its ranking accuracy on queries that were never matched during training; the paper’s hard-negative mining assumes such unmatched queries can still carry competitive mask evidence, and that assumption is only empirically supported.
  • CLSD’s best-teacher selection could be swapped for an exponential-moving-average teacher of the final layer to isolate whether the improvement comes from depth diversity or from a smoother distillation target; the paper does not run this comparison.
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 / 4 minor

Summary. The paper proposes iFAN, a training-only framework for plain mask transformers, consisting of two objectives: Adjusted Probability-Mask Ranking (APMR) and Cross-Layer Self-Distillation (CLSD). APMR learns a global mask-quality score, rescales the probability-mask score map, and ranks matched queries above hard unmatched competitors. CLSD selects a strong intermediate-layer teacher query by soft-IoU and distills its adjusted score map to the final layer. The method keeps inference as final-layer-only decoding. Experiments on COCO, ADE20K, and Cityscapes with PMT and EoMT baselines report consistent PQ, AP, and mIoU gains, with headline averages of 1.20 PQ, 1.30 AP, and 0.63 mIoU and no additional inference cost.

Significance. If the results hold, iFAN is a practically valuable, architecture-agnostic training strategy that improves segmentation accuracy without changing the inference procedure. The paper's strengths include paired comparisons against matched baselines, a modular ablation study (Tables 4 and 5), and a diagnostic analysis (Figure 4 and Table 9) that directly measures the proposed mechanisms. The training objectives are not circular: the quality head is supervised by soft-IoU from ground truth, ranking gradients are stopped on class and mask branches, and CLSD teacher selection uses soft-IoU; the gains are therefore empirical rather than by construction. However, the headline AP average is not reproducible from the paper's own corrected table, and the efficiency claim is based on accounting rather than direct measurement.

major comments (3)
  1. [Table 3 / Appendix Table 8 / Abstract] The appendix explicitly states: 'For the PMT ViT-L 1280 results reported in the main paper, there was a transcription error. The correct results are the ones provided in the appendix.' Comparing Table 3 with Table 8: the main text reports the PMT ViT-L 1280 baseline as 48.8 AP and iFAN as 50.6 AP (+1.8); Table 8 gives the baseline as 49.3 AP, so the true gain is +1.3 AP. With this correction, the four AP deltas are 1.2, 1.3, 1.2, and 1.8, averaging 1.375, not the abstract's 1.30 AP. The main text phrase 'improves PMT by an average of 1.1 AP' is also inconsistent with both versions of Table 3. This load-bearing error must be fixed in the table, the abstract, and the main-text summary.
  2. [Tables 1-3 footnotes; Abstract/Conclusion] The paper claims 'negligible additional parameters, FLOPs and inference latency' (Abstract) and 'virtually no additional inference overhead' (Main Results), but the tables footnote that iFAN rows follow the matched baseline under 'shared-configuration accounting rather than an independent efficiency measurement.' The added quality head (a single linear layer plus sigmoid) and the adjusted scoring operation ((q_i * p_i)^gamma * M_i) are not directly measured. Please either report actual parameter counts, FLOPs, and latencies for iFAN and baselines, or soften the claims to expected/estimated overhead. The current wording overstates what the evidence supports.
  3. [Experimental Setup; Tables 2 and 5(b)] Several reported gains are small (e.g., +0.3 mIoU for EoMT on Cityscapes in Table 2, +0.8 PQ for ViT-G in Table 5(b)) and all are single-validation numbers without error bars or multiple seeds. Because the central claim is 'consistent improvements' across settings, the absence of variance estimates makes it difficult to assess whether the smallest gains are within run-to-run noise. Please provide error bars or multiple-run results for the key comparisons, or explicitly note the single-run limitation.
minor comments (4)
  1. [Adjusted Probability-Mask Ranking, Eq. (3)-(4)] The definition of the unmatched-query score map has a typo: 'denoting by the probability-mask score map of unmatched query S_j = ((q_j * p_j(c_k))^gamma, M_j)' should use multiplication (i.e., S_j = (q_j * p_j(c_k))^gamma * M_j) to be consistent with Eq. (4).
  2. [Tables 1-3] The footnote about 'shared-configuration accounting' appears in all three main tables; consider consolidating it into the experimental setup or using a single notation explanation to reduce redundancy.
  3. [Supplementary Material, first paragraph] 'Tables 6-8 We provide more detailed results' is missing a colon or period after 'Tables 6-8'.
  4. [Figure 4 (right)] The stacked bar labels (53, 80 and 98, 35) are not immediately clear; please clarify in the caption which segments correspond to intermediate vs. final layers for baseline and iFAN.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified; the reported gains are empirical training-objective improvements, not reductions to fitted targets or self-citations.

full rationale

The paper's derivation chain is self-contained and empirical. APMR learns a mask-quality score q_i with BCE supervision against soft-IoU of the matched mask (Eq. 3-4), and the ranking loss updates only the quality branch via stop-gradient, leaving class and mask branches intact; no reported PQ/AP/mIoU number is set equal to a fitted target by construction. CLSD selects an intermediate teacher by argmax soft-IoU (Eq. 5) and distills its adjusted score map with a one-sided hinge (Eq. 6), but this is a training objective, not a prediction that reduces to its input. Hyperparameters lambda_rank=0.1 and lambda_distill=0.4 are chosen from sensitivity sweeps (Fig. 3), which is standard tuning rather than circularity. The paper cites prior work (Mask DINO, DN-DETR, Rank-DETR) only as related context; no load-bearing premise is justified solely by a self-citation or an imported uniqueness theorem. Two transparency issues are noted but are not circularity: the Table 1 footnote states that 'Parameters, GFLOPs, and FPS in iFAN rows follow the matched baseline under shared-configuration accounting rather than an independent efficiency measurement,' and the appendix states 'For the PMT ViT-L 1280 results reported in the main paper, there was a transcription error. The correct results are the ones provided in the appendix.' These affect the reliability of the efficiency and AP-average claims but do not constitute a reduction of the derivation to its own inputs.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a set of hyperparameters selected by validation sweeps and on standard domain assumptions about mask-quality supervision and the fixed inference rule. No new physical or conceptual entities are introduced; the quality head is a learned component rather than an invented mechanism.

free parameters (5)
  • lambda_rank = 0.1
    Ranking loss weight chosen by sweep (Fig. 3); robust across 0.025 to 0.20.
  • lambda_distill = 0.4
    Distillation loss weight chosen by sweep (Fig. 3); robust across 0.05 to 0.80.
  • N_H hard negatives = 10
    Number of hard negative unmatched queries per target; stable for 7 to 15 (Fig. 5).
  • gamma = 2
    Exponent in adjusted probability-mask score; varies less than 0.2 PQ for gamma in [0.5, 3] (Fig. 5).
  • CLSD schedule = start epoch 4, linear annealing
    Warm-up and annealing of lambda_distill; mentioned in implementation, not swept in figures.
assumptions (3)
  • domain assumption Soft-IoU of the predicted mask with the ground-truth mask is a valid per-query training target for mask-quality estimation
    Used in APMR quality-head BCE loss and in CLSD teacher selection (Eq. 5); if soft-IoU is a poor proxy, the calibrated scores and selected teachers would be unreliable.
  • domain assumption The inference assembly rule of pixel-wise argmax over probability-mask scores is fixed and not modified
    Motivates training-time APMR; the assumption is that improving the training signal under this fixed rule transfers to better final outputs.
  • domain assumption Intermediate layers of the trained mask transformer can be used as teachers during training without changing the inference architecture
    CLSD relies on the soft-IoU of intermediate Hungarian-matched queries and the shared quality head across layers; the paper provides empirical support but no guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of iFAN: Inference-Aware Learning for Plain Mask Transformers." pith.science (2026). https://pith.science/paper/JSCDJQFM

@misc{pith2026260803216,
  author       = {Pith},
  title        = {Pith review of: iFAN: Inference-Aware Learning for Plain Mask Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JSCDJQFM}},
  note         = {Machine review of arXiv:2608.03216}
}
read the original abstract

Query-based mask transformers assemble segmentation outputs through pixel-wise competition among query predictions of the final layer, yet this inference process is not explicitly optimized during training. We identify two key mismatches: the query with the highest probability-mask score does not necessarily produce the most accurate mask, and final-layer decoding may discard superior predictions from intermediate layers. To address these issues, we propose Inference-Aware Learning (iFAN), a general training framework for plain mask transformers. iFAN introduces Adjusted Probability-Mask Ranking (APMR), which aligns query competition with predicted mask quality and suppresses high-confidence but inaccurate competitors. We further employ Cross-Layer Self-Distillation (CLSD) to transfer stronger intermediate predictions to the final layer. The ranking and distillation objectives are training-only, while inference retains efficient final-layer decoding. Experiments on COCO, ADE20K, and Cityscapes demonstrate consistent improvements across panoptic, instance, and semantic segmentation, as well as across different architectures, backbone scales, and input resolutions. Overall, iFAN improves performance by an average of 1.20 PQ, 1.30 AP, and 0.63 mIoU, with negligible additional parameters, FLOPs and inference latency.

Figures

Figures reproduced from arXiv: 2608.03216 by the authors.

Figure 1
Figure 1. We study the performance of mask transformer on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of iFAN. APMR addresses the query-wise mismatch by learning mask quality, calibrating the scores used for assembly, and ranking a matched query above high-evidence unmatched query. CLSD addresses the layer-wise mismatch by selecting a stronger teacher query from an intermediate layer and transferring its assembled evidence to the final layer. At inference, the adjusted-prob-mask is only computed for final l… view at source ↗
Figure 3
Figure 3. evaluates the sensitivity of iFAN with EoMT-L and PMT-L to the loss weights of the APMR and CLSD ob￾jectives. Both models remain robust across a broad range of hyperparameter values. We therefore adopt representa￾tive values within the stable regions, namely λrank = 0.10 and λdistill = 0.40, and keep them fixed across all datasets, 0.025 0.05 0.10 0.15 0.20 0.40 0.80 λrank 56.4 56.5 56.6 56.7 56.8 56.9 57.0 PQ 56.76… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Inference diagnostics on COCO val2017. Left: disagreement between the max-prob-mask and max-IoU queries, together with the frequency of high probability-mask scores paired with low IoU. Right: the numbers of classes whose optimal prediction is produced by an intermedia…
Figure 5
Figure 5. Figure 5: evaluates the sensitivity to the number of hard neg￾atives NH and the weighting exponent γ. Performance re￾mains consistently strong for NH ∈ [7, 15]. Similarly, vary￾ing γ from 0.5 to 3 results in only a 0.18 PQ difference. The setting γ = 2 effectively suppresses low…
Figure 6
Figure 6. Figure 6: Qualitative results for panoptic segmentation. Each row shows an input image, ground truth, and the standalone prediction. The examples contain complex street layouts with large stuff regions and multiple thing instances, including vehicles, pedestrians, traffic signs,…
Figure 7
Figure 7. Figure 7: Qualitative results for instance segmentation. The three columns show the input image, the ground truth, and the standalone instance masks. The selected scenes emphasize dense arrangements, substantial occlusion, scale variation, and adjacent instances with similar app…
Figure 8
Figure 8. Figure 8: Qualitative results for semantic segmentation. Each row presents the input image, the ground truth, and the standalone prediction. The examples cover diverse indoor and outdoor scenes and illustrate coherent region assignment around objects with different shapes, scale…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 20 canonical work pages

  1. [1]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Dn-detr: Accelerate detr training by introducing query denoising , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

    work page 2023

  2. [2]

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

    Detection transformer with stable matching , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  3. [3]

    The eleventh international conference on learning representations , year=

    Dino: Detr with improved denoising anchor boxes for end-to-end object detection , author=. The eleventh international conference on learning representations , year=

    work page

  4. [4]

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

    Panoptic segformer: Delving deeper into panoptic segmentation with transformers , 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=

    Mp-former: Mask-piloted transformer for image segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  6. [6]

    arXiv preprint arXiv:2010.11929 , year=

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv preprint arXiv:2010.11929 , year=

    arXiv 2010

  7. [7]

    European conference on computer vision , pages=

    End-to-end object detection with transformers , author=. European conference on computer vision , pages=. 2020 , organization=

    2020

  8. [8]

    Advances in neural information processing systems , volume=

    Per-pixel classification is not all you need for semantic segmentation , author=. Advances in neural information processing systems , volume=

Show all 41 references
  1. [9]

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

    Masked-attention mask transformer for universal image segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  2. [10]

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

    Oneformer: One transformer to rule universal image segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  3. [11]

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

    Mask dino: Towards a unified transformer-based framework for object detection and segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  4. [12]

    arXiv preprint arXiv:2205.08534 , year=

    Vision transformer adapter for dense predictions , author=. arXiv preprint arXiv:2205.08534 , year=

    arXiv

  5. [13]

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

    Swin transformer: Hierarchical vision transformer using shifted windows , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  6. [14]

    European conference on computer vision , pages=

    k-means Mask Transformer , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  7. [15]

    arXiv preprint arXiv:2209.15001 , year=

    Dilated neighborhood attention transformer , author=. arXiv preprint arXiv:2209.15001 , year=

    arXiv

  8. [16]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    2009

  9. [17]

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

    Depth anything: Unleashing the power of large-scale unlabeled data , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  10. [18]

    Transactions on Machine Learning Research Journal , year=

    Dinov2: Learning robust visual features without supervision , author=. Transactions on Machine Learning Research Journal , year=

  11. [19]

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

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

  12. [20]

    European conference on computer vision , pages=

    Microsoft coco: Common objects in context , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  13. [21]

    Proceedings of the computer vision and pattern recognition conference , pages=

    Your vit is secretly an image segmentation model , author=. Proceedings of the computer vision and pattern recognition conference , pages=

  14. [22]

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

    Mask scoring r-cnn , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  15. [23]

    Advances in neural information processing systems , volume=

    Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection , author=. Advances in neural information processing systems , volume=

  16. [24]

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

    Varifocalnet: An iou-aware dense object detector , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  17. [25]

    , author=

    Align-DETR: Enhancing End-to-end Object Detection with Aligned Loss. , author=. BMVC , year=

  18. [26]

    Advances in Neural Information Processing Systems , volume=

    Rank-DETR for high quality object detection , author=. Advances in Neural Information Processing Systems , volume=

  19. [27]

    Artificial intelligence and statistics , pages=

    Deeply-supervised nets , author=. Artificial intelligence and statistics , pages=. 2015 , organization=

    2015

  20. [28]

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

    Be your own teacher: Improve the performance of convolutional neural networks via self distillation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  21. [29]

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

    Detrdistill: A universal knowledge distillation framework for detr-families , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  22. [30]

    arXiv preprint arXiv:2406.05791 , year=

    OD-DETR: online distillation for stabilizing training of detection transformer , author=. arXiv preprint arXiv:2406.05791 , year=

    arXiv

  23. [31]

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

    Scene parsing through ade20k dataset , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  24. [32]

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

    The cityscapes dataset for semantic urban scene understanding , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  25. [33]

    International journal of computer vision , volume=

    The pascal visual object classes (voc) challenge , author=. International journal of computer vision , volume=. 2010 , publisher=

    2010

  26. [34]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    arXiv

  27. [35]

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

    Max-deeplab: End-to-end panoptic segmentation with mask transformers , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  28. [36]

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

    Mask transfiner for high-quality instance segmentation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  29. [37]

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

    Pseudo-label alignment for semi-supervised instance segmentation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  30. [38]

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

    Enhanced training of query-based object detection via selective query recollection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  31. [39]

    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=

  32. [40]

    arXiv preprint arXiv:2603.25398 , year=

    PMT: Plain Mask Transformer for Image and Video Segmentation with Frozen Vision Encoders , author=. arXiv preprint arXiv:2603.25398 , year=

  33. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Shahabodini, Sajjad and Mansoori, Mobina and Bayatmakou, Farnoush and Abouei, Jamshid and Plataniotis, Konstantinos and Mohammadi, Arash , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2026 , pages =

    2026

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.