Pith. sign in

REVIEW 3 major objections 5 minor 47 references

MOCHA: Multi-modal Objects-aware Cross-arcHitecture Alignment

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read MOCHA distills multimodal region-level knowledge from a frozen vision-language model into a lightweight detector, improving few-shot personalized object detection by +10.1 average over the YOLOv8n baseline.

desk verdict A credible, practical distillation recipe for few-shot personalised detection; the headline gain is probably real in direction but partly tuned to the four benchmarks, and the loss weights are missing. read the letter →

arxiv 2509.14001 v5 pith:2MDWIO54 submitted 2025-09-17 cs.CV cs.AIcs.LG

Elena Camuffo , Francesco Barbato , Mete Ozay , Simone Milani , Umberto Michieli This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords personalizedobjectdetectionfew-shotlearningknowledgedistillationvision-languagemodelcross-architecturerelationalembeddinglossPCAcompressionlightweightdetector
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

The paper claims that a compact, vision-only object detector can acquire the fine-grained semantic discrimination needed for few-shot personalization by distilling region-level embeddings from a frozen vision-language model. The teacher's fused visual-textual embeddings are PCA-compressed and channel-normalized, then used to train a lightweight student with a dual loss: pointwise alignment (L1+L2) plus a relational cross-entropy that preserves the global geometry of the teacher's embedding space. On four personal-detection benchmarks (PerSeg, POD, CORe50, iCubWorld) under 1- and 5-shot regimes, MOCHA reports an average +10.1 improvement over an undestilled YOLOv8n and +4.9 over the closest prior method (AuXFT), at a cost of about 3 ms per image. The method requires no teacher and no text at inference; only a frozen student backbone and a prototype classifier are used.

What carries the argument

The load-bearing object is the teacher target u'_i: a PCA-compressed (512-d) concatenation of the CLIP class token and the LLaVa language-model output for each region, rescaled per channel by a fitted power law σc ≈ 18/(c+1)^0.47 − 0.26. Around this, the method builds (i) a translation module tS—a transformer encoder block with channel-wise multi-head self-attention plus MLP—that maps the student's concatenated multi-scale pooled region features into the teacher space; (ii) a pointwise distillation loss Ldist averaging L1 and L2 distances; and (iii) a relational embedding loss Lemb that converts pairwise Euclidean distances into softmax distributions and minimises their cross-entropy. Togeth

What would settle it

Run the MOCHA pipeline unchanged except that the PCA subspace and the channel-normalization curve are fit on a dataset from a domain far from OpenImages (e.g., medical or aerial imagery) while keeping the rest of the distillation identical; if the +10-point gain over the undestilled YOLOv8n baseline disappears while a variant that refits PCA on the target domain retains it, the transferability of the teacher-target subspace is the load-bearing premise.

Watch

Extended reading notes

Core claim

MOCHA's central claim is that a lightweight detector can be taught the fine-grained semantic structure of a large multimodal model by distilling, not the VLM's raw outputs, but a PCA-compressed, channel-normalized version of its fused visual-textual region embeddings. The paper shows that combining a pointwise distillation loss with a relational embedding loss—which matches pairwise distance distributions between student and teacher features—gives the student both accurate local alignment and global structural consistency in the teacher's embedding space. The result is that the student's features become separable enough that a simple nearest-class-mean classifier can recognize a user's speci

Load-bearing premise

The PCA axes and per-channel power-law normalization fitted on the OpenImages distillation set must transfer to the personal domain; if a user's objects lie outside that learned linear subspace, the student is trained toward a miscalibrated target space and the reported gains could shrink or reverse.

Editorial extensions

If this is right

  • A vision-only YOLO-sized detector can reach near-VLM-level semantic granularity for personalization, enabling on-device few-shot recognition without a text encoder at inference.
  • The distilled representations are architecture-agnostic: the same supervision improves YOLOv8n, YOLOv11n/s/l, and RT-DETR-l students, and works with ProtoNet, SimpleShot, and linear-probing classifiers.
  • PCA compression of the 4608-d multimodal embedding to 512-d not only cuts memory/compute but in the paper's experiments improves or maintains accuracy across datasets, implying much of the relevant semantics lives in a low-dimensional subspace.
  • Multimodal supervision (visual + textual) outperforms either modality alone, so language descriptions add complementary discriminative signal even for a vision-only student.
  • Because teacher targets can be cached offline, distillation adds only ~3 ms/image at inference and, when starting from AuXFT weights, only ~11 h of training, making the recipe practical for resource-constrained deployment.

Reading between the lines

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

  • If the teacher's multimodal embedding space is indeed a shared semantic space, the same distillation recipe—PCA-compressed fused embeddings plus relational loss—could transfer other grounding capabilities (e.g., referring expressions, attribute recognition) into compact detectors by changing only the supervision dataset and re-fitting the PCA/normalization.
  • The relational loss, which explicitly preserves pairwise distance geometry, attacks the neural-collapse failure mode that the paper cites for low-data detectors; a testable extension is whether this loss alone, without the pointwise term, can regularize self-supervised pretraining of small models on scarce data.
  • The empirical power-law for channel standard deviations suggests teacher embeddings have a highly skewed energy profile; if this shape is generic, one could replace the per-dataset PCA fit with a closed-form whitening transform, simplifying deployment to new domains.
  • Since the teacher runs only on the distillation dataset and its outputs are cached, the same framework could distill from an even larger VLM or ensemble without increasing the student's inference footprint—only the offline cache and the PCA fit would change.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper presents MOCHA, a three-stage distillation framework for few-shot personalized object detection. A frozen LLaVa-1.5-7B teacher produces region-level multimodal embeddings (visual CLIP token plus text-conditioned LLaVa representation), which are PCA-compressed, channel-normalized, and used as targets for a lightweight YOLOv8n student through a translation module. The training objective combines a pointwise L1+L2 distillation loss with a pairwise-distance relational embedding loss. After distillation, the student backbone and translator are frozen and a prototype classifier is trained on 1/5 support examples. Experiments on PerSeg, POD, CORe50, and iCubWorld report average mAP/mAcc improvements of +10.1 over YOLOv8n and +4.9 over AuXFT, with ablations across student architectures, FSL classifiers, PCA dimensions, and pretraining. The appendix includes Wilcoxon tests, computational cost comparisons, pseudocode, and a limitations section.

Significance. If the empirical claims survive scrutiny, MOCHA is a practically relevant contribution: it transfers multimodal knowledge offline while keeping inference cheap, and the systematic ablations across architectures and FSL classifiers, plus the released code, are valuable. The relational embedding loss is a sensible addition over pointwise distillation. However, the central quantitative claim is currently weakened by the absence of a validation-based hyperparameter protocol and missing loss weights, so the magnitude of the reported gains is not yet firmly established.

major comments (3)
  1. [Sec. 4.2, Fig. 4; Sec. 4.4, Tab. 4] The PCA dimension dt=512 is selected by measuring mAP on the same four personal benchmarks used for the main results. Fig. 4 reports an 'average score across all evaluation datasets' and Tab. 4 explicitly selects 512 as the best balance; neither a held-out validation set nor a separate model-selection protocol is described. Since the main claim in Tab. 2 (avg 58.7 vs 53.8 for AuXFT) is produced with this dt, the reported gains are vulnerable to test-set overfitting. Please either select dt on a validation split disjoint from the test benchmarks, or report results for a fixed default dt and for a range of dt values without selection.
  2. [Sec. 3.4, Eq. (9)] The final objective L = L_det + lambda_dist * L_dist + lambda_emb * L_emb is never instantiated: the paper reports no values for lambda_dist and lambda_emb, and Tab. 4 does not ablate them. Without these values the method is not reproducible, and if they were tuned on the benchmarks the risk of overfitting compounds the issue in the previous comment. Please report the values used in all experiments and add an ablation (or at least a sensitivity table) over lambda_dist/lambda_emb.
  3. [Appendix A.3.3, Eq. (3), Fig. 7] All main results use channel-normalized targets u'_i, where sigma_c is replaced by a fitted power law 18/(c+1)^0.47 - 0.26. The paper shows only the fitted curve; it gives no residual/error measure, no number of channels over which the fit is valid, and no evidence that this normalization transfers from OpenImages to the personal datasets. Since this fitted curve directly defines the teacher targets for L_dist and L_emb, an inaccurate fit could distort the student's target space and inflate or deflate the reported gains. Please quantify the fit error and test sensitivity to the normalization (e.g., using the empirical per-channel sigma_c instead of the fitted curve).
minor comments (5)
  1. [A.2] Title contains a typo: 'Perfomance Evaluation' should be 'Performance Evaluation'.
  2. [Sec. 4.3, paragraph after Tab. 2] The text says the retrieval protocol [3] allows multiple candidate boxes, but it is not clear how this interacts with mAP for PerSeg/POD in Tab. 2. Please define the metric precisely so the reader does not have to consult [3] to interpret the scores.
  3. [Algorithm 1] The pseudocode uses K both as batch size and as a sampled batch ('Sample K ∼ D_c, with |K|=K'). Rename one of them to avoid notation clash.
  4. [Tables 9–11] The caption says mAP/mAcc, but mAcc is not defined in the main text. Please define it and specify which columns correspond to which metric.
  5. [Sec. 4.1, Datasets] Please specify the exact class labels C_c used to prompt the teacher during distillation (e.g., OpenImages class names) and state whether the same labels are used for all datasets. This is needed for reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

PCA dimension d_t is selected on the evaluation benchmarks, so part of the reported gain is fitted rather than predicted.

  1. fitted input called prediction [Section 4.2, 'Impact of Teacher Feature Dimensionality' (Fig. 4); selection reiterated in Section 4.4 'Components Design']
    "Fig. 4 evaluates the role of teacher supervision dimensionality (dt) by comparing features: (i) of the personal datasets Df (as an upper bound), (ii) of the distillation dataset Dc unnormalized and (iii) PCA normalized. Overall, performance increases steadily up to dt = 512, which strikes a good balance between compactness and expressivity."

    Fig. 4 is captioned 'Average score across all evaluation datasets varying feature dimension d_t.' The evaluation datasets are PerSeg, POD, CORe50, and iCubWorld—the same four benchmarks used for the headline Tab. 2 results and the '+10.1 average improvement' claim. The paper then adopts d_t=512 as 'the embedding size' and reports the main results at that value. Choosing a free hyperparameter by averaging mAP on the test benchmarks, and then reporting those same benchmarks as MOCHA's out-of-sample performance, makes the reported gain partially a selected value rather than a prediction. The unreported λ_dist and λ_emb (Eq. 9) further prevent checking whether other free parameters were also tuned on these benchmarks.

full rationale

The core distillation mechanism is not circular: the teacher is a frozen externally-pretrained LLaVa model, the student is evaluated on personal datasets unseen during distillation, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the result. The self-citation to AuXFT is used as a baseline and initialization, but the MOCHA distillation itself is an independent contribution. The main circular step is the selection of PCA dimension d_t on the evaluation data. Fig. 4 explicitly computes average mAP across the four personal benchmarks as a function of d_t, and Section 4.4 then adopts d_t=512 before reporting the headline +10.1 improvement on exactly those benchmarks. This is the 'fitted input called prediction' pattern: a free hyperparameter is fit to the test data, and the resulting prediction is presented as if it were an out-of-sample result. Because the method still outperforms AuXFT at other d_t values in Tab. 4, the circularity is partial rather than total; the reported quantitative gain, however, is partly constructed by the test-set selection. The missing loss weights in Eq. (9) are a reproducibility risk but not by themselves evidence of circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The method rests on external teacher features and standard learning tools; no new physical or mathematical entities are introduced. The main assumptions are domain-specific transferability of the teacher's embedding space and a fitted normalization curve used to rescale teacher targets.

free parameters (4)
  • PCA dimension d_t = 512
    Selected based on benchmark mAP over dimensions (Fig 4); balances compactness and expressiveness; tuned on the evaluation datasets.
  • Power-law coefficients (a, b, c) for channel normalization = a=18, b=0.47, c=-0.26
    Fitted to channel-wise std of teacher PCA embeddings on OpenImages (Fig 7); used in Eq. 3 for per-channel rescaling.
  • Distillation loss weight lambda_dist = not reported
    Weight in the final objective Eq. 9; paper does not give values or sensitivity analysis.
  • Embedding loss weight lambda_emb = not reported
    Weight in the final objective Eq. 9; paper does not give values or sensitivity analysis.
assumptions (6)
  • domain assumption Teacher (LLaVa) multimodal embeddings, after PCA on OpenImages and channel rescaling, are a transferable supervisory signal for personal object detection.
    The entire distillation objective trains the student to mimic u'_i computed from OpenImages; personal datasets are never used for distillation, so the OpenImages-derived PCA and scaling must transfer across domains. Invoked in Section 3.2 and evaluated in Tabs 9-11.
  • domain assumption Semantically related concepts exhibit similar embedding structure across modalities in a well-generalized backbone.
    Motivates why aligning student visual features to teacher multimodal targets should transfer semantics; introduced in Section 1.
  • domain assumption Averaging the LLM output tokens yields a meaningful region descriptor h_i.
    Section 3.2: 'We average the output along the temporal dimension to obtain a single semantic embedding h_i'; no analysis of token aggregation choice.
  • domain assumption Concatenating the CLIP class token z_V,i (scaled by gamma = ||h_i||) with h_i preserves both appearance and semantics in a fixed dimension.
    Section 3.2 Eq. 1; the gamma scale is chosen by hand, and the concatenated space is assumed to be a faithful joint representation.
  • domain assumption Pairwise Euclidean distance structure in the teacher's target space is the right relational signal for student regularization.
    Section 3.4, Lemb in Eqs. 7-8; assumes distance-based neighborhood preservation improves generalization for few-shot prototype classification.
  • ad hoc to paper PCA is a valid dimensionality reduction for teacher embeddings, and the fitted power-law sigma_c curve approximates per-channel scales accurately enough.
    Appendix A.3.3: sigma_c is fitted to observed channel stds; used in Eq. 3 to normalize targets. If the fit is inaccurate, teacher targets are miscalibrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MOCHA: Multi-modal Objects-aware Cross-arcHitecture Alignment." pith.science (2026). https://pith.science/paper/2MDWIO54

@misc{pith2026250914001,
  author       = {Pith},
  title        = {Pith review of: MOCHA: Multi-modal Objects-aware Cross-arcHitecture Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2MDWIO54}},
  note         = {Machine review of arXiv:2509.14001}
}
read the original abstract

Personalized object detection aims to adapt a general-purpose detector to recognize user-specific instances from only a few examples. Lightweight models often struggle in this setting due to their weak semantic priors, while large vision-language models (VLMs) offer strong object-level understanding but are too computationally demanding for real-time or on-device applications. We introduce MOCHA (Multi-modal Objects-aware Cross-arcHitecture Alignment), a distillation framework that transfers multimodal region-level knowledge from a frozen VLM teacher into a lightweight vision-only detector. MOCHA extracts fused visual and textual teacher's embeddings and uses them to guide student training through a dual-objective loss that enforces accurate local alignment and global relational consistency across regions. This process enables efficient transfer of semantics without the need for teacher modifications or textual input at inference. MOCHA consistently outperforms prior baselines across four personalized detection benchmarks under strict few-shot regimes, yielding a +10.1 average improvement, with minimal inference cost.

Figures

Figures reproduced from arXiv: 2509.14001 by the authors.

Figure 1
Figure 1. MOCHA recipe. (1) Pretraining student model. (2) Knowledge distillation on rich joint visual and textual features from a frozen teacher. (3) Few-shot personalization with frozen student and prototypical learner. criminative power. In this work, we aim to bridge this gap by proposing MOCHA, a knowledge distillation approach that transfers object-centric multimodal embeddings from a vision-language teacher into a comp… view at source ↗
Figure 2
Figure 2. MOCHA system. (a) Feature distillation: Student detector is trained on dataset Dc aligning multiscale region-level features to PCA-pruned multimodal embeddings from a frozen vision-language teacher via the translation module tS(·). (b) Personalization: Student backbone and Translator are frozen and used to compute semantic prototypical features from a personal dataset Df . These are then used to train a prototype-ba… view at source ↗
Figure 3
Figure 3. Effect of Lemb on a set of ten 2D points optimized with respect to 3D ones. (a) 3D reference points, proxy for the teacher embeddings u ′ i . (b) Evolution of the 2D points (proxy for student embeddings f ′ A,i) updated via Lemb from the 3D reference points u ′ i . ⋆ marks the original location, timesteps increase with color saturation. (c) Percent rate of 2D top-k nearest neighbors (k-NN) that match those of the re… view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Feature similarity between the FA embeddings and the teacher target u ′ i encoded in the R/G/B channels, one for each object in the input scene from POD dataset. pretrained weights (which in turn start from COCO and fine￾tune on OpenImages). Distillation is performed o…
Figure 6
Figure 6. Figure 6: Qualitative results on PerSeg. Ground truth in green, prediction with the highest confidence in red (class names shown refer to personal class labels in PerSeg). YOLOv8n refers to base￾line. *: refers to MOCHA (AuXFT). 4.3. Main Results We begin our evaluation in the o…
Figure 7
Figure 7. Figure 7: Fitting the PCA curve σ ≃ a (x+1)b + c, a = 18, b = 0.47, c = −0.26. Model PerSeg POD CORe50 iCubWorld Avg 1 SHOT 1 SHOT 5 SHOT 1 SHOT 5 SHOT 1 SHOT 5 SHOT LLaVa (ui) (4608 channels) 94.0 ±1.8 66.4 ±3.8 80.7 ±0.0 59.4 ±4.6 80.4 ±4.3 58.0 ±3.2 80.6 ±2.4 74.2 PCA @ 1024 …
Figure 8
Figure 8. Figure 8: Qualitative results in the 1-shot setting. Ground truth in green, prediction with the highest confidence in red (class names shown [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results in the 5-shot setting. Ground truth in green, prediction with the highest confidence in red (class names shown [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 2 linked inside Pith

  1. [1]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Men- sch, Katie Millicah, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Shar- ifzadeh, Mikolaj Binkow...

    2022

  2. [2]

    Continual road-scene semantic segmen- tation via feature-aligned symmetric multi-modal network

    Francesco Barbato, Elena Camuffo, Simone Milani, and Pietro Zanuttigh. Continual road-scene semantic segmen- tation via feature-aligned symmetric multi-modal network. In IEEE International Conference on Image Processing (ICIP), pages 722–728. IEEE, 2024. 2

    2024

  3. [3]

    Cross-architecture auxiliary fea- ture space translation for efficient few-shot personalized ob- ject detection

    Francesco Barbato, Umberto Michieli, Jijoong Moon, Pietro Zanuttigh, and Mete Ozay. Cross-architecture auxiliary fea- ture space translation for efficient few-shot personalized ob- ject detection. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024. 2, 3, 4, 5, 7, 9

    2024

  4. [4]

    Learn- ing from mistakes: Self-regularizing hierarchical representa- tions in point cloud semantic segmentation.IEEE Transac- tions on Multimedia, 2023

    Elena Camuffo, Umberto Michieli, and Simone Milani. Learn- ing from mistakes: Self-regularizing hierarchical representa- tions in point cloud semantic segmentation.IEEE Transac- tions on Multimedia, 2023. 1

    2023

  5. [5]

    Bert: Pre-training of deep bidirectional transform- ers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transform- ers for language understanding. InNorth American Chapter of the Association for Computational Linguistics, 2019. 2

    2019

  6. [6]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 2

    arXiv 2010

  7. [7]

    iCub World: Friendly Robots Help Building Good Vision Data-Sets

    Sean Ryan Fanello, Carlo Ciliberto, Matteo Santoro, Lorenzo Natale, Giorgio Metta, Lorenzo Rosasco, and Francesca Odone. iCub World: Friendly Robots Help Building Good Vision Data-Sets. InIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 700–705, 2013. 5

    2013

  8. [8]

    Imagebind: One embedding space to bind them all

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15180–15190, 2023. 1

    2023

Show all 47 references
  1. [9]

    Reciprocal teacher-student learning via forward and feedback knowledge distillation

    Jianping Gou, Yu Chen, Baosheng Yu, Jinhua Liu, Lan Du, Shaohua Wan, and Zhang Yi. Reciprocal teacher-student learning via forward and feedback knowledge distillation. IEEE Transactions on Multimedia, 26:7901–7916, 2024. 2

    2024

  2. [10]

    Vild: Open-vocabulary object detection via vision and lan- guage knowledge distillation.International Conference on learning Representations (ICLR), 2022

    Jialin Gu, Golnaz Ghiasi, Yin Cui, Zhonghua Wang, et al. Vild: Open-vocabulary object detection via vision and lan- guage knowledge distillation.International Conference on learning Representations (ICLR), 2022. 2, 5

    2022

  3. [11]

    CDFKD-MFS: Collaborative Data-Free Knowledge Distilla- tion via Multi-Level Feature Sharing.IEEE Transactions on Multimedia, 24:4262–4274, 2022

    Zhiwei Hao, Yong Luo, Zhi Wang, Han Hu, and Jianping An. CDFKD-MFS: Collaborative Data-Free Knowledge Distilla- tion via Multi-Level Feature Sharing.IEEE Transactions on Multimedia, 24:4262–4274, 2022. 2

    2022

  4. [12]

    One-for-all: Bridge the gap be- tween heterogeneous architectures in knowledge distillation

    Zhiwei Hao, Jianyuan Guo, Kai Han, Yehui Tang, Han Hu, Yunhe Wang, and Chang Xu. One-for-all: Bridge the gap be- tween heterogeneous architectures in knowledge distillation. InIEEE International Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 5

    2023

  5. [13]

    The platonic representation hypothesis.Proceedings of Machine Learning Research (PMLR), 2024

    Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis.Proceedings of Machine Learning Research (PMLR), 2024. 1

    2024

  6. [14]

    xmuda: Cross-modal unsu- pervised domain adaptation for 3d semantic segmentation

    Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Em- ilie Wirbel, and Patrick P ´erez. xmuda: Cross-modal unsu- pervised domain adaptation for 3d semantic segmentation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12605–12614, 2020. 2

    2020

  7. [15]

    Cross-modal learning for domain adaptation in 3d semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(2): 1533–1544, 2022

    Maximilian Jaritz, Tuan-Hung Vu, Raoul De Charette, ´Emilie Wirbel, and Patrick P´erez. Cross-modal learning for domain adaptation in 3d semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(2): 1533–1544, 2022. 2

    2022

  8. [16]

    Ultralytics yolo11, 2024

    Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 1, 5, 9

    2024

  9. [17]

    Ultralytics yolov8 [computer software]

    Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8 [computer software]. https://github.com/ultralytics/ultralytics. accessed july 2024, 2023. 1, 2, 9

    2024

  10. [18]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 4015–4026. IEEE, 2023. 2

    2023

  11. [19]

    Neural collapse: A review on mod- elling principles and generalization.Transactions on Machine Learning Research, 2022

    Vignesh Kothapalli. Neural collapse: A review on mod- elling principles and generalization.Transactions on Machine Learning Research, 2022. 1

    2022

  12. [20]

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open im- ages dataset v4: Unified image classification, object detection, and visual relationship detection...

    1956

  13. [21]

    Lightweight model pre- training via language guided knowledge distillation.IEEE Transactions on Multimedia, 26:10720–10730, 2024

    Mingsheng Li, Lin Zhang, Mingzhen Zhu, Zilong Huang, Gang Yu, Jiayuan Fan, and Tao Chen. Lightweight model pre- training via language guided knowledge distillation.IEEE Transactions on Multimedia, 26:10720–10730, 2024. 2

    2024

  14. [22]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740–755. Springer, 2014. 5

    2014

  15. [23]

    Visual instruction tuning.Advances in Neural Information Processing Systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in Neural Information Processing Systems, 36:34892–34916, 2023. 1, 2, 3

    2023

  16. [24]

    Cross-architecture knowledge distillation

    Yufan Liu, Jiajiong Cao, Bing Li, Weiming Hu, Jingting Ding, and Liang Li. Cross-architecture knowledge distillation. In IEEE/CVF Proceedings of the Asian Conference on Computer Vision (ACCV), pages 3396–3411, 2022. 2

    2022

  17. [25]

    Matcher: Segment anything with one shot using all-purpose feature matching.International Conference on Learning Representations (ICLR), 2023

    Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching.International Conference on Learning Representations (ICLR), 2023. 2 14

    2023

  18. [26]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012–10022. IEEE, 2021. 2

    2021

  19. [27]

    Core50: a new dataset and benchmark for continuous object recognition,

    Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition,

  20. [28]

    Knowledge amalgamation from hetero- geneous networks by common feature learning.International Joint Conference on Artificial Intelligence (IJCAI), 2019

    Sihui Luo, Xinchao Wang, Gongfan Fang, Yao Hu, Dapeng Tao, and Mingli Song. Knowledge amalgamation from hetero- geneous networks by common feature learning.International Joint Conference on Artificial Intelligence (IJCAI), 2019. 2

    2019

  21. [29]

    Rtdetrv2: All-in-one detection transformer beats yolo and dino, 2024

    Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, and Yi Liu. Rtdetrv2: All-in-one detection transformer beats yolo and dino, 2024. 5

    2024

  22. [30]

    Toward founda- tion models for inclusive object detection: Geometry- and category-aware feature extraction across road user categories

    Zonglin Meng, Xin Xia, and Jiaqi Ma. Toward founda- tion models for inclusive object detection: Geometry- and category-aware feature extraction across road user categories. IEEE Transactions on Systems, Man, and Cybernetics: Sys- tems, 54(11):6570–6580, 2024. 2

    2024

  23. [31]

    Object-conditioned bag of instances for few-shot personalized instance recognition

    Umberto Michieli, Jijoong Moon, Daehyun Kim, and Mete Ozay. Object-conditioned bag of instances for few-shot personalized instance recognition. InIEEE/SPS Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 7885–7889. IEEE, 2024. 2

    2024

  24. [32]

    Maxime Oquab, Timoth´ee Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicola...

    2023 arXiv

  25. [33]

    Vardan Papyan, X. Y . 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. 1

    2020

  26. [34]

    Swiss dino: Efficient and versatile vision framework for on-device personal object search

    Kirill Paramonov, Jia-Xing Zhong, Umberto Michieli, Jijoong Moon, and Mete Ozay. Swiss dino: Efficient and versatile vision framework for on-device personal object search. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2564–2571. IEEE, 2024. 2

    2024

  27. [35]

    Het- erogeneous knowledge distillation using information flow modeling

    Nikolaos Passalis, Maria Tzelepi, and Anastasios Tefas. Het- erogeneous knowledge distillation using information flow modeling. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2339–2348. IEEE, 2020. 2

    2020

  28. [36]

    Learning transferable visual models from natural language supervision.Proceedings of Machine Learning Research (PMLR), 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision.Proceedings of Machin...

    2021

  29. [37]

    Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Y . Bengio. Fitnets: Hints for thin deep nets. InInternational Conference on Learned Representations (ICLR), 2015. 2, 5

    2015

  30. [38]

    Prototypi- cal networks for few-shot learning

    Jake Snell, Kevin Swersky, and Richard Zemel. Prototypi- cal networks for few-shot learning. InAdvances in Neural Information Processing Systems, 2017. 1, 3, 9

    2017

  31. [39]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through at- tention. InInternational Conference on Machine Learning (ICML), pages 10347–10357, 2021. 5

    2021

  32. [40]

    Hybrid knowledge distillation network for RGB-D co- salient object detection.IEEE Transactions on Systems, Man, and Cybernetics: Systems, pages 1–12, 2025

    Zhangping Tu, Wujie Zhou, Xiaohong Qian, and Weiqing Yan. Hybrid knowledge distillation network for RGB-D co- salient object detection.IEEE Transactions on Systems, Man, and Cybernetics: Systems, pages 1–12, 2025. 2

    2025

  33. [41]

    Attention is all you need.Advances in Neural Information Processing Systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in Neural Information Processing Systems, 30, 2017. 2

    2017

  34. [42]

    Seggpt: Segmenting everything in context.IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chun- hua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context.IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2

    2023

  35. [43]

    Weinberger, and Laurens van der Maaten

    Yan Wang, Wei-Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning, 2019. 9

    2019

  36. [44]

    Cmda: Cross-modality domain adap- tation for nighttime semantic segmentation

    Ruihao Xia, Chaoqiang Zhao, Meng Zheng, Ziyan Wu, Qiyu Sun, and Yang Tang. Cmda: Cross-modality domain adap- tation for nighttime semantic segmentation. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 21572–21581, 2023. 2

    2023

  37. [45]

    Task-oriented feature distillation.Advances in Neural Information Processing Systems, 33:14759–14771,

    Linfeng Zhang, Yukang Shi, Zuoqiang Shi, Kaisheng Ma, and Chenglong Bao. Task-oriented feature distillation.Advances in Neural Information Processing Systems, 33:14759–14771,

  38. [46]

    Personalize segment anything model with one shot

    Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junt- ing Pan, Hao Dong, Yu Qiao, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. InIn- ternational Conference on Learning Representations (ICLR),

  39. [47]

    Personalized image semantic segmen- tation

    Yu Zhang, Chang-Bin Zhang, Peng-Tao Jiang, Ming-Ming Cheng, and Feng Mao. Personalized image semantic segmen- tation. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 10549–10559. IEEE, 2021. 2 15

    2021

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.