Pith. sign in

REVIEW 5 major objections 4 minor 53 references

Temporally Consistent Unsupervised Segmentation for Mobile Robot Perception

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Frontier-Seg claims that unsupervised terrain segmentation for mobile robots can be made temporally consistent—without labels, tracking, or motion estimation—by clustering foundation-model superpixel features locally in short windows and…

desk verdict A clearly built label-free segmentation pipeline with a plausible temporal-consistency story, but the headline temporal metric is partly guaranteed by construction and the paper never measures per-frame stability directly. read the letter →

arxiv 2507.22194 v1 pith:NJLRJW5Y submitted 2025-07-29 cs.CV cs.RO

classification cs.CVcs.RO
keywords unsupervisedsegmentationtemporalconsistencyterrainmobilerobotperceptionDINOv2superpixelclusteringoff-roadnavigationvideo
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

Frontier-Seg tries to show that a mobile robot can segment terrain into coherent, temporally stable regions from raw video with no human labels and no motion estimation or tracking. The method pools self-supervised foundation-model features (DINOv2) inside SLIC superpixels, clusters the resulting region descriptors within short temporal windows to form pseudo-labels, merges the superpixels, then re-clusters the merged descriptors across the full sequence to fix a global label vocabulary. On the off-road RUGD and RELLIS-3D benchmarks this two-phase design more than doubles temporal mIoU relative to the zero-shot baseline DiffCut (34.15 vs 13.94 on RUGD; 31.12 vs 13.79 on RELLIS-3D) while also reducing under-segmentation entropy. This matters because robots deployed in unstructured, unfamiliar environments cannot rely on fixed semantic taxonomies or labor-intensive ground-truth annotation, and inconsistent frame-by-frame segmentation undermines navigation and planning.

What carries the argument

The load-bearing mechanism is two-phase clustering with a region-descriptor recomputation step. First, region descriptors $z_{t,m}$ come from masked average pooling of DINOv2 dense features over SLIC superpixels. Local K-means ($K=100$) is applied to descriptors aggregated over non-overlapping windows of 100 frames, producing pseudo-labels $\ell(z)$ that merge superpixels into regions; those merged regions yield recomputed descriptors $\hat{z}_{t,m'}$ by masked average pooling. Global K-means ($K=50$) over all recomputed descriptors assigns the final labels $\lambda(\hat{z})$, and the label of each merged region is propagated to its pixels. The recomputation is what lets the features be re-pooled at a coarser spatial scale, aligning them with the evolving region structure before the global assignment.

What would settle it

Recompute the temporal evaluation using a per-frame cluster-to-class mapping or an adjacent-frame label-flip metric (for instance, the temporal consistency metric cited as [14]) on the same RUGD and RELLIS-3D outputs; if Frontier-Seg's temporal mIoU and accuracy advantage over DiffCut shrinks to near zero under that metric, the claim that global clustering delivers temporal consistency is not supported.

Watch

Extended reading notes

Core claim

The central discovery is that temporal consistency can be obtained from clustering alone, without motion cues. Region descriptors are formed by masked average pooling of dense features inside superpixels; local K-means over 100-frame windows assigns pseudo-labels that merge superpixels; then the merged descriptors are recomputed and globally clustered (K=50) across the whole video to assign final labels. The paper argues that because the final clustering is global, every frame shares one label vocabulary, and it reports substantially higher temporal mIoU and pixel accuracy than applying the same temporal pipeline to DiffCut's zero-shot outputs, with lower under-segmentation entropy on RUGD and comparable stability on RELLIS-3D. The authors present this as a label-free foundation for terrain perception that adapts to new environments without retraining.

Load-bearing premise

The load-bearing premise is that temporal consistency is fairly measured by matching each predicted cluster to a ground-truth class once over the entire video sequence, rather than per frame—since the method assigns all final labels in one global clustering pass, a fixed label set across frames is guaranteed by construction, and a per-frame label-stability metric could shrink the reported advantage over the baseline.

Editorial extensions

If this is right

  • Off-road robots can obtain terrain segmentations with temporally stable labels from raw video alone, so entering a new environment requires no annotation campaign and no predefined class list.
  • Because the cluster count K sets the ontology granularity, an operator can choose fine or coarse pseudo-classes without retraining the feature backbone.
  • DINOv2 features cost roughly 250 ms per 512×512 frame versus about 1.2 s for DiffCut's diffusion-based inference, so the temporal gains come at lower compute and are better suited to resource-constrained robots.
  • The method needs no optical flow or tracking, so it applies where motion estimation is unreliable, such as occluded, texture-poor, or high-speed off-road scenes.
  • The reported results establish a quantitative baseline for temporally consistent unsupervised segmentation on RUGD and RELLIS-3D, with temporal mIoU of 34.15 and 31.12 respectively.

Reading between the lines

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

  • A natural extension the paper leaves implicit is an online variant that replaces the single global K-means pass with incremental centroid updates over sliding windows, lifting the offline limitation stated in the Limitations section while preserving label stability.
  • The DINOv2-versus-SSD-1B performance gap suggests that the quality of self-supervised dense features is the main lever; testing other self-supervised vision transformers under the same two-phase pipeline would show whether the temporal gains transfer.
  • Because the method produces stable pseudo-labels, those labels could serve as weak supervision to train a lightweight student segmentation network for the same environment, compounding the label-free advantage.
  • An ablation that removes the local-window stage and runs only the global clustering would isolate how much of the temporal consistency comes from the windowing step versus the global assignment.
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

5 major / 4 minor

Summary. The paper proposes Frontier-Seg, an unsupervised, temporally consistent terrain segmentation pipeline for mobile-robot video. The method extracts DINOv2 per-pixel features, pools them into SLIC superpixel region descriptors, performs K-means clustering within non-overlapping temporal windows, merges regions by local pseudo-label, recomputes descriptors, and finally runs a global K-means over the full sequence to assign one label vocabulary across all frames. The authors evaluate on RUGD and RELLIS-3D against DiffCut, reporting strong temporal mIoU/accuracy gains (e.g., 34.15 vs 13.94 temporal mIoU on RUGD) while acknowledging lower zero-shot accuracy in several settings. The main claim is that global re-clustering of temporally aggregated foundation-model features yields temporally consistent pseudo-labels without motion estimation or supervision.

Significance. If the temporal-consistency claim were directly supported, the paper would offer a practical recipe for unsupervised terrain segmentation in off-road robotics, with a clearly described two-level clustering pipeline and an honest limitations section. The strengths include a transparent pipeline description, a re-evaluation of a strong zero-shot baseline using the authors' code, and extensive per-subdataset supplementary tables that expose cluster-count sensitivity. The central contribution, however, rests on two measurements that are not yet demonstrated: (i) a fair like-for-like temporal comparison with the unmodified baseline, and (ii) a direct per-frame label-stability metric. The paper's own cited metric for exactly this purpose (Varghese et al. [14]) is never computed. As written, the headline 34.15 vs 13.94 temporal mIoU gap is consistent with a pipeline-internal artifact rather than a measured property of temporal consistency, so the significance is currently conditional on an added experiment.

major comments (5)
  1. [Sec. 4.1, Table 1] The zero-shot comparison is not like-for-like. DiffCut is evaluated per frame, while Frontier-Seg in the zero-shot setting runs the full temporal model (local plus global clustering over the entire sequence) and only applies Hungarian matching per frame. This conflates temporal aggregation with the zero-shot capability being compared. Please report a per-frame variant of Frontier-Seg that omits global clustering (and ideally also local temporal windows, or explicitly ablate each), so the zero-shot numbers reflect the same information available to DiffCut.
  2. [Sec. 3.5, Sec. 4.2, Supp. C.1] The temporal mIoU and accuracy numbers in Table 1 are computed with a single cluster-to-class mapping over the whole video sequence. Since Frontier-Seg's global K-means (Eq. 4) assigns one label vocabulary for all frames by construction, a fixed global mapping is guaranteed, and the reported temporal advantage over DiffCut may be partly built into the pipeline rather than measured as label stability. The paper cites Varghese et al. [14] but never computes that unsupervised temporal consistency metric. Please add a direct per-frame stability measurement, such as adjacent-frame label flip rates or the Varghese et al. metric, computed before any global cluster-to-class alignment, for both Frontier-Seg and DiffCut. This is the load-bearing evidence for the paper's central claim.
  3. [Sec. 4.2 vs Supp. C.1] The main text states that many-to-one Hungarian matching is used for cluster-to-class alignment before computing mIoU and accuracy, while Supplementary C.1 (Eq. 10) explicitly replaces Hungarian matching with majority voting and even notes that the approach differs from Hungarian-based methods. This is a direct internal inconsistency in the evaluation protocol. Please state which alignment is used for each table and make the main text and supplementary agree; otherwise the reported metrics cannot be reproduced.
  4. [Sec. 4.1, Table 1] The DiffCut temporal numbers are not produced by the unmodified DiffCut algorithm: the authors feed DiffCut's zero-shot segmentation outputs and SSD-1B features into their own local/global clustering pipeline. The comparison is therefore Frontier-Seg versus a DiffCut+Frontier-Seg hybrid, not Frontier-Seg versus DiffCut in the temporal setting. The headline temporal mIoU gap (34.15 vs 13.94 on RUGD) cannot be attributed to Frontier-Seg's superiority over DiffCut's intrinsic temporal behavior. Please report unmodified DiffCut temporal numbers (with whatever alignment is used for all methods) or explicitly rename the baseline as an ablation of the clustering stage.
  5. [Sec. 4.3, Tables 1 and 2-15] No statistical significance or run-to-run variability is reported for any K-means-based result. Since K-means initialization is stochastic and the evaluations cover only five to eighteen sequences, the reported differences (e.g., the zero-shot mIoU gap of 56.00 vs 53.93 on RUGD) could fall within random variation. Please provide standard deviations across multiple random restarts or otherwise justify that the reported differences are stable.
minor comments (4)
  1. [Sec. 3.4] There are typos in Sec. 3.4 and Sec. 1: "pseudo-labels" is written as "psuedo-labels" in the introduction's contribution list and "pseudo-labels" is misspelled in the local-clustering description; please correct them.
  2. [Sec. 4.1/Figure legend] The qualitative figure (Fig. 2) shows only single frames; for a paper about temporal consistency, a qualitative side-by-side of the same terrain region across consecutive frames (with label stability indicated) would make the visual claim more concrete.
  3. [Supp. C.1] Equation (10) uses K for predicted clusters in the text but the preceding paragraph defines K as the number of clusters while C is the number of classes; the equation is clear, but the surrounding notation in Supp. C.1 would benefit from a consistent statement that the mapping is many-to-one and not one-to-one.
  4. [Sec. 2] The related-work section mentions streaming-data segmentation [47] as similar in using local and global clustering, but does not say precisely how Frontier-Seg differs from that work in terms of features or downstream evaluation; a sentence contrasting the feature backbones and the evaluation setting would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: method is an unsupervised clustering cascade evaluated against held-out ground truth; construction-based temporal metric is an evaluation concern, not a derived prediction.

full rationale

Frontier-Seg's pipeline (Secs. 3.3-3.5, Eqs. 1-4) is a self-contained clustering cascade: DINOv2/SLIC region descriptors, local K-means pseudo-labeling, region merging, descriptor recomputation, then global K-means for final labels. No parameter is fitted to ground truth; RUGD/RELLIS-3D annotations are used only for evaluation (Sec. 4.1). The temporal-consistency claim rests on Eq. 4, which assigns all final labels in one global K-means pass, and the evaluation then applies one cluster-to-class mapping over the whole sequence (Supp. C.1). This makes a fixed label vocabulary across frames a property of the construction, but the comparison is not circular: DiffCut is run through the same local/global clustering pipeline (Sec. 4.1), so the reported temporal gap reflects differences in features and region proposals rather than a hidden reuse of the target result. The Sec. 4.2 description of 'single Hungarian matching' is inconsistent with the majority-voting rule defined in Supp. C.1, and the paper cites the per-frame temporal-consistency metric of Varghese et al. [14] without computing it; these are evaluation-validity concerns, not circularity. Self-citations (e.g., [47] for local/global clustering and [19] in related work) are contextual and are not load-bearing for the main derivation. No equation reduces to its own input by construction, and no fitted parameter is renamed as a prediction. Therefore the derivation is self-contained with no significant circularity.

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

The central claim rests on several domain assumptions: DINOv2 features encode terrain semantics, SLIC superpixels align with terrain boundaries, Euclidean K-means groups terrain into meaningful classes, and smooth egomotion holds in the test datasets. The method introduces no new physical entities. The main free parameters are the number of local and global clusters, window length, and superpixel size, all chosen by hand; K_global is the most consequential because it fixes the ontology size yet is not selected on an independent validation set.

free parameters (5)
  • K_local = 100
    Number of clusters for local K-means pseudo-label assignment; chosen without a stated validation criterion and controls over-segmentation.
  • K_global = 50
    Number of final terrain categories in the main results; supplementary sweeps K from 12 to 400, but the main value is not justified by an independent validation set.
  • temporal_window_size = 100 frames (Delta t = 99)
    Non-overlapping window length for local clustering; chosen by hand, with no reported sensitivity analysis.
  • superpixel_size = SLIC region size 30, about 200 superpixels per frame
    Initial segmentation granularity; chosen by hand and affects the quality of pooled descriptors.
  • gaussian_blur_sigma = 0.7
    Preprocessing smoothing in CIELAB space; chosen by hand.
assumptions (5)
  • domain assumption DINOv2 dense features are semantically meaningful for terrain without fine-tuning
    Section 3.3 relies on DINOv2 features being rich enough that averaged superpixel descriptors separate terrain classes.
  • domain assumption Euclidean K-means over averaged DINOv2 features yields terrain-relevant clusters
    Sections 3.4 and 3.5 assume K-means in feature space groups regions into meaningful, consistent terrain categories.
  • domain assumption SLIC superpixels follow terrain boundaries well enough that pooled descriptors are discriminative
    Section 3.3 uses SLIC as the initial segmentation; if superpixels straddle terrain boundaries, the pooled descriptors are corrupted.
  • domain assumption Smooth frame-to-frame continuity and consistent egomotion hold in the test sequences
    Section 6 explicitly states this assumption and notes it may break down in dynamic scenes or rapid viewpoint changes.
  • domain assumption Ground-truth annotations are used only for evaluation and not for clustering decisions
    Section 4.1 states ground truth is never seen during training; this excludes the possibility that hyperparameters were selected using test labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Temporally Consistent Unsupervised Segmentation for Mobile Robot Perception." pith.science (2026). https://pith.science/paper/NJLRJW5Y

@misc{pith2026250722194,
  author       = {Pith},
  title        = {Pith review of: Temporally Consistent Unsupervised Segmentation for Mobile Robot Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NJLRJW5Y}},
  note         = {Machine review of arXiv:2507.22194}
}
read the original abstract

Rapid progress in terrain-aware autonomous ground navigation has been driven by advances in supervised semantic segmentation. However, these methods rely on costly data collection and labor-intensive ground truth labeling to train deep models. Furthermore, autonomous systems are increasingly deployed in unrehearsed, unstructured environments where no labeled data exists and semantic categories may be ambiguous or domain-specific. Recent zero-shot approaches to unsupervised segmentation have shown promise in such settings but typically operate on individual frames, lacking temporal consistency-a critical property for robust perception in unstructured environments. To address this gap we introduce Frontier-Seg, a method for temporally consistent unsupervised segmentation of terrain from mobile robot video streams. Frontier-Seg clusters superpixel-level features extracted from foundation model backbones-specifically DINOv2-and enforces temporal consistency across frames to identify persistent terrain boundaries or frontiers without human supervision. We evaluate Frontier-Seg on a diverse set of benchmark datasets-including RUGD and RELLIS-3D-demonstrating its ability to perform unsupervised segmentation across unstructured off-road environments.

Figures

Figures reproduced from arXiv: 2507.22194 by the authors.

Figure 1
Figure 1. Overview of the Frontier-Seg pipeline. Local clustering aggregates region descriptors over [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison on four RUGD sequences. Columns show the input frame, ground [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 46 canonical work pages

  1. [14]

    Varghese, Y

    S. Varghese, Y . Bayzidi, A. Bar, N. Kapoor, S. Lahiri, J. D. Schneider, N. M. Schmidt, P. Schlicht, F. Huger, and T. Fingscheidt. Unsupervised temporal consistency metric for video segmentation in highly-automated driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 336–337, 2020

  2. [1]

    Garcia-Garcia, S

    A. Garcia-Garcia, S. Orts-Escolano, S. Oprea, V . Villena-Martinez, P. Martinez-Gonzalez, and J. Garcia-Rodriguez. A survey on deep learning techniques for image and video semantic segmentation.Applied Soft Computing, 70:41–65, 2018

  3. [2]

    Lateef and Y

    F. Lateef and Y . Ruichek. Survey on semantic segmentation using deep learning techniques. Neurocomputing, 338:321–348, 2019

  4. [3]

    Van Brummelen, M

    J. Van Brummelen, M. O’Brien, D. Gruyer, and H. Najjaran. Autonomous vehicle perception: The technology of today and tomorrow.Transportation research part C: emerging technolo- gies, 89:384–406, 2018

  5. [4]

    R. R. Murphy.Disaster robotics. MIT press, 2014

  6. [5]

    Nagatani, S

    K. Nagatani, S. Kiribayashi, Y . Okada, K. Otake, K. Yoshida, S. Tadokoro, T. Nishimura, T. Yoshida, E. Koyanagi, M. Fukushima, et al. Emergency response to the nuclear accident at the fukushima daiichi nuclear power plants using mobile rescue robots.Journal of Field Robotics, 30(1):44–63, 2013

  7. [6]

    Reina, A

    G. Reina, A. Milella, and R. Galati. Terrain assessment for precision agriculture using vehicle dynamic modelling.Biosystems engineering, 162:124–139, 2017

  8. [7]

    La Hera, O

    P. La Hera, O. Mendoza-Trejo, O. Lindroos, H. Lideskog, T. Lindb ¨ack, S. Latif, S. Li, and M. Karlberg. Exploring the feasibility of autonomous forestry operations: Results from the first experimental unmanned machine.Journal of Field Robotics, 41(4):942–965, 2024

Show all 53 references
  1. [8]

    Szab ´o and Z

    L. Szab ´o and Z. Weltsch. A comprehensive review of existing datasets for off-road autonomous vehicles. In2024 IEEE 22nd World Symposium on Applied Machine Intelligence and Infor- matics (SAMI), pages 000403–000410. IEEE, 2024. 9

  2. [9]

    D. Feng, C. Haase-Sch ¨utz, L. Rosenbaum, H. Hertlein, C. Glaeser, F. Timm, W. Wiesbeck, and K. Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges.IEEE Transactions on Intelligent Transportation Syst...

  3. [10]

    J. Guo, U. Kurup, and M. Shah. Is it safe to drive? an overview of factors, metrics, and datasets for driveability assessment in autonomous driving.IEEE Transactions on Intelligent Transportation Systems, 21(8):3135–3151, 2019

  4. [11]

    Couairon, M

    P. Couairon, M. Shukor, J.-E. HAUGEARD, M. Cord, and N. THOME. Diffcut: Catalyz- ing zero-shot semantic segmentation with diffusion features and recursive normalized cut. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openrev...

  5. [12]

    X. Wang, R. Girdhar, S. X. Yu, and I. Misra. Cut and learn for unsupervised object detection and instance segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3124–3134, 2023

  6. [13]

    Shuai, Y

    Z. Shuai, Y . Chen, S. Mao, Y . Zho, and X. Zhang. Diffseg: a segmentation model for skin lesions based on diffusion difference.arXiv preprint arXiv:2404.16474, 2024

  7. [15]

    Wigness, S

    M. Wigness, S. Eum, J. G. Rogers, D. Han, and H. Kwon. A rugd dataset for autonomous navigation and visual perception in unstructured outdoor environments. In2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5000–5007. IEEE, 2019

  8. [16]

    Jiang, P

    P. Jiang, P. Osteen, M. Wigness, and S. Saripalli. Rellis-3d dataset: Data, benchmarks and analysis. In2021 IEEE international conference on robotics and automation (ICRA), pages 1110–1116. IEEE, 2021

  9. [17]

    X. Meng, N. Hatch, A. Lambert, A. Li, N. Wagener, M. Schmittle, J. Lee, W. Yuan, Z. Chen, S. Deng, et al. Terrainnet: Visual modeling of complex terrain for high-speed, off-road navi- gation.arXiv preprint arXiv:2303.15771, 2023

  10. [18]

    Miyamoto, Y

    R. Miyamoto, Y . Nakamura, M. Adachi, T. Nakajima, H. Ishida, K. Kojima, R. Aoki, T. Oki, and S. Kobayashi. Vision-based road-following using results of semantic segmentation for autonomous navigation. In2019 IEEE 9th International Conference on Consumer Electronics (ICCE-Berl...

  11. [19]

    Wigness, J

    M. Wigness, J. G. R. III, C.-E. Tsai, C. Mertz, L. Navarro-Serment, and J. Oh. Using perception cues for context-aware navigation in dynamic outdoor environments.Field Robotics, 1(1):1 – 33, October 2021

  12. [20]

    Ellis, M

    C. Ellis, M. Wigness, J. Rogers, C. Lennon, and L. Fiondella. Risk averse bayesian reward learning for autonomous navigation from human demonstration. In2021 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS), pages 8928–8935. IEEE, 2021

  13. [21]

    Y . Tan, N. Virani, B. Good, S. Gray, M. Yousefhussien, Z. Yang, K. Angeliu, N. Abate, and S. Sen. Risk-aware autonomous navigation. InArtificial Intelligence and Machine Learning for Multi-Domain Operations Applications III, volume 11746, pages 335–348. SPIE, 2021

  14. [22]

    M. Liu, E. Yurtsever, J. Fossaert, X. Zhou, W. Zimmer, Y . Cui, B. L. Zagar, and A. C. Knoll. A survey on autonomous driving datasets: Statistics, annotation quality, and a future outlook. IEEE Transactions on Intelligent Vehicles, 2024. 10

  15. [23]

    Parmar, S

    J. Parmar, S. Chouhan, V . Raychoudhury, and S. Rathore. Open-world machine learning: applications, challenges, and opportunities.ACM Computing Surveys, 55(10):1–37, 2023

  16. [24]

    Shi and J

    J. Shi and J. Malik. Normalized cuts and image segmentation.IEEE Transactions on pattern analysis and machine intelligence, 22(8):888–905, 2000

  17. [25]

    Comaniciu and P

    D. Comaniciu and P. Meer. Mean shift: A robust approach toward feature space analysis.IEEE Transactions on pattern analysis and machine intelligence, 24(5):603–619, 2002

  18. [26]

    P. F. Felzenszwalb and D. P. Huttenlocher. Efficient graph-based image segmentation.Inter- national journal of computer vision, 59:167–181, 2004

  19. [27]

    T. F. Chan and L. A. Vese. Active contours without edges.IEEE Transactions on image processing, 10(2):266–277, 2001

  20. [28]

    Boykov and V

    Y . Boykov and V . Kolmogorov. An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision.IEEE transactions on pattern analysis and machine intelli- gence, 26(9):1124–1137, 2004

  21. [29]

    Arbelaez, M

    P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik. Contour detection and hierarchical image segmentation.IEEE transactions on pattern analysis and machine intelligence, 33(5):898– 916, 2010

  22. [30]

    Achanta, A

    R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. S¨usstrunk. Slic superpixels compared to state-of-the-art superpixel methods.IEEE transactions on pattern analysis and machine intelligence, 34(11):2274–2282, 2012

  23. [31]

    Caron, P

    M. Caron, P. Bojanowski, A. Joulin, and M. Douze. Deep clustering for unsupervised learning of visual features. InProceedings of the European conference on computer vision (ECCV), pages 132–149, 2018

  24. [32]

    X. Ji, J. F. Henriques, and A. Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865–9874, 2019

  25. [33]

    W.-C. Hung, V . Jampani, S. Liu, P. Molchanov, M.-H. Yang, and J. Kautz. Scops: Self- supervised co-part segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 869–878, 2019

  26. [34]

    J. H. Cho, U. Mall, K. Bala, and B. Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794–16804, 2021

  27. [35]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. HAZIZA, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut,...

  28. [36]

    J. Zhou, C. Wei, H. Wang, W. Shen, C. Xie, A. Yuille, and T. Kong. ibot: Image bert pre- training with online tokenizer.arXiv preprint arXiv:2111.07832, 2021

  29. [37]

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000–16009, 2022

  30. [38]

    Hamilton, Z

    M. Hamilton, Z. Zhang, B. Hariharan, N. Snavely, and W. T. Freeman. Unsupervised semantic segmentation by distilling feature correspondences.arXiv preprint arXiv:2203.08414, 2022. 11

  31. [39]

    Weinzaepfel, J

    P. Weinzaepfel, J. Revaud, Z. Harchaoui, and C. Schmid. Deepflow: Large displacement optical flow with deep matching. InProceedings of the IEEE international conference on computer vision, pages 1385–1392, 2013

  32. [40]

    Revaud, P

    J. Revaud, P. Weinzaepfel, Z. Harchaoui, and C. Schmid. Epicflow: Edge-preserving interpo- lation of correspondences for optical flow. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1164–1172, 2015

  33. [41]

    E. Ilg, N. Mayer, T. Saikia, M. Keuper, A. Dosovitskiy, and T. Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2462–2470, 2017

  34. [42]

    X. Zhu, Y . Xiong, J. Dai, L. Yuan, and Y . Wei. Deep feature flow for video recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2349– 2358, 2017

  35. [43]

    J. Su, R. Yin, S. Zhang, and J. Luo. Motion-state alignment for video semantic segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3571–3580, 2023

  36. [44]

    H. Wang, W. Wang, and J. Liu. Temporal memory attention for video semantic segmentation. In2021 IEEE International Conference on Image Processing (ICIP), pages 2254–2258. IEEE, 2021

  37. [45]

    M. Gao, F. Zheng, J. J. Yu, C. Shan, G. Ding, and J. Han. Deep learning for video object segmentation: a review.Artificial Intelligence Review, 56(1):457–531, 2023

  38. [46]

    C. Xu, C. Xiong, and J. J. Corso. Streaming hierarchical video segmentation. InEuropean Conference on Computer Vision, pages 626–639. Springer, 2012

  39. [47]

    Wigness and J

    M. Wigness and J. G. Rogers. Unsupervised semantic scene labeling for streaming data. InPro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4612– 4621, 2017

  40. [48]

    Darcet, M

    T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski. Vision transformers need registers.arXiv preprint arXiv:2309.16588, 2023

  41. [49]

    Gupta, V

    Y . Gupta, V . V . Jaddipal, H. Prabhala, S. Paul, and P. V on Platen. Progressive knowledge distillation of stable diffusion xl using layer level loss.arXiv preprint arXiv:2401.02677, 2024

  42. [50]

    H. W. Kuhn. The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2):83–97, 1955

  43. [51]

    Gong and J

    H. Gong and J. Shi. Conditional entropies as over-segmentation and under-segmentation met- rics for multi-part image segmentation. university of pennsylvania department of computer and information science; philadelphia, pa. Technical report, USA: 2011. Technical Report MS-CIS-...

  44. [52]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman. The pascal visual object classes (voc) challenge.International journal of computer vision, 88(2):303–338, 2010

  45. [53]

    Cordts, M

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele. The cityscapes dataset for semantic urban scene understanding. InProceedings of the Conference on Computer Vision and Pattern Recognition, pages 3213–3223. IEEE, 2016. 12 ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.