Pith. sign in

REVIEW 2 major objections 4 minor 33 references

SR-JEPA: Learning Predictive Latent State in 3D Scenes

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A frozen JEPA predictor, queried at the location of a deleted object, outputs a latent that identifies the missing object's semantic class, and this content combines with geometry for structural decisions.

desk verdict A genuinely controlled study of what a frozen JEPA predictor completes in 3D, with a live room-prior confound and an underpowered structural endpoint. read the letter →

arxiv 2608.05774 v1 pith:N4GQVCOU submitted 2026-08-06 cs.CV cs.LG

classification cs.CVcs.LG
keywords self-supervisedlearningjoint-embeddingpredictivearchitecture3Dsceneunderstandingpointcloudlatentpredictionobjectcompletionsemanticidentityspatialrelations
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

SR-JEPA asks what a joint-embedding predictive architecture infers when an entire object is removed from a native 3D scene, and answers that the frozen predictive pathway completes a latent state carrying the missing object's semantic identity. Training uses only self-contained 3D EMA targets from 1,201 ScanNet scenes: no reconstruction, labels, language, or 2D foundation features. At evaluation, every point of one object is deleted and replaced by a fixed shape-free 32-point query at its centroid. On 5,953 held-out ARKitScenes objects the imputed latent reaches 43.13% semantic-identity macro accuracy, 22.18 points above the strongest shortcut floor; randomizing the prediction path removes 9.78 points and substituting a matched donor context removes 21.98 points. On 8,570 Sr3D support pairs, identity decoded from the imputed latent combined with anchor identity and geometry reaches 39.37 AP against 41.15 AP for the full latent, indicating that completed entity content and metric geometry compose into a structural decision.

What carries the argument

The mechanism that carries the argument is the deletion-first fixed-query protocol combined with a deliberately narrow, retained predictor. The context encoder (a Point Transformer V3) processes visible occupied cells; an EMA target encoder observes the full scene and supplies regression targets; and the predictor, conditioned on visible state and query locations, outputs a 1232-dimensional latent per query point. At evaluation, all points of one object are removed before voxelization and a fixed 32-point sphere of radius 0.15 m is inserted at the oracle centroid, so the query carries no class, shape, or orientation information. The trained-versus-randomized-predictor graft and correct-versus-donor context contrasts isolate what the learned prediction path and the correct scene actually contribute, while joint identity–geometry controls test whether downstream support utility goes beyond completed identity and metric geometry.

What would settle it

Replace the true surrounding scene with a matched donor scene that has the same token count and extent but different furniture; the paper reports accuracy drops from 32.94% to 10.95%. If a donor scene matched additionally on room function, such as kitchen-to-kitchen, recovers high accuracy, then the context effect is a room-layout prior rather than entity completion.

Watch

Extended reading notes

Core claim

The paper's central claim is that a trained JEPA predictor, kept frozen, is a queryable and compositional 3D predictive state: when an entire entity is absent, the predictor completes entity content that depends on the visible surrounding scene, and that content can be combined with metric geometry by a downstream computation. The evidence is the deletion-first protocol, in which the target object's points are removed before encoding, a fixed shape-free query is placed at the supplied centroid, and the predictor output is read out with MLP probes. The trained predictor beats the strongest shortcut floor by 22.18 points, loses 9.78 points when its weights are randomized under a fixed encoder, and loses 21.98 points when the correct scene is replaced by a matched donor scene. The support endpoint then shows that predicted target identity plus anchor identity and coordinates reproduces most of the full latent's utility, leaving an unresolved residual that is not consistently positive across checkpoints. The authors state the result as a division of labor: the predictive pathway estimates what belongs at the queried location, and a downstream function combines that entity state with where to determine support.

Load-bearing premise

The central claim relies on treating a supervised MLP probe trained on frozen latents as a faithful instrument, so that probe decodability of the predicted latent is evidence of completed semantic content rather than of room-location priors or deletion-boundary and donor-geometry shortcuts.

Editorial extensions

If this is right

  • The retained predictor can serve as an operational state interface for present-state inference under partial observability, rather than being discarded after pretraining.
  • An entirely self-contained 3D JEPA objective, without reconstruction or external supervision, can produce latents that carry semantic identity of fully absent objects.
  • Downstream structural reasoning (support detection) can be implemented as a lightweight function over completed identity plus coordinates, with the full-latent residual left as an open question.
  • Identity completion is robust across two training recipes, reaching 43.13% and 44.61% macro accuracy, so the phenomenon is not tied to one loss-weighting choice.
  • The correct scene context is the dominant driver: replacing it with a matched donor reduces identity accuracy by 21.98 points, so the predictive state is contextual, not a location prior alone.

Reading between the lines

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

  • A natural extension the paper does not test is to use the same frozen predictor for object discovery: query many candidate locations, and treat regions where the completed latent has high self-consistency or low prediction error as evidence of a missing entity.
  • Because the compositional control still leaves a small, sign-inconsistent residual on support, a finer-grained endpoint such as orientation, size, or contact geometry might reveal what, if anything, the full latent adds beyond identity and coordinates.
  • The evaluation principle—report what the predictor completes, not just what the encoder recognizes—could transfer directly to video JEPAs and LiDAR world models, where the predictor is already trained but rarely probed after deletion of whole entities.
  • The reliance on MLP probes suggests a cheap complementary falsification: nearest-neighbor retrieval in the frozen latent space, without any trained head, should rank the true class highly if the latent genuinely encodes identity.
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

2 major / 4 minor

Summary. The paper introduces SR-JEPA, a point-native joint-embedding predictive architecture for 3D scenes. It pretrains a sparse scene encoder and a narrow predictor on ScanNet using only EMA latent targets, then evaluates the frozen predictor directly by deleting all points of an object, inserting a fixed shape-free 32-point query at its centroid, and probing the resulting imputed latent. On ARKitScenes, the imputed latent yields 43.13% macro identity accuracy, with a +22.18-point gap over the strongest floor, a +9.78-point drop when the predictor is randomized, and a +21.98-point drop when correct context is replaced by a matched donor. On Sr3D, the same latent supports a composition analysis in which predicted identity plus anchor identity and geometry reaches 39.37 AP against 41.15 AP for the full latent. The paper concludes that the predictive pathway completes context-dependent entity content that a downstream computation combines with metric geometry.

Significance. If the claim holds, this is a valuable and unusually rigorous demonstration that the frozen predictor of a JEPA is a testable scientific object. The deletion-first protocol, the union-hole donor design, scene-disjoint splits, multiple checkpoints and training recipes, paired scene-cluster bootstrap intervals, and explicit artifact provenance are strong methodological commitments. The identity result is replicated across two training recipes, and the paper is unusually candid about its limitations. The main reservation is whether the imputed latent encodes room-level or spatial-layout priors rather than local, object-specific completion; this is testable with same-room-type donors or room-stratified analyses and should be resolved before the compositional interpretation is fully accepted.

major comments (2)
  1. [§4.2, Appendix C, Table 2] The donor-context contrast is the load-bearing control for the claim that completion depends on the correct surrounding scene, but the donors are matched only by visible-token count and bounding-box diagonal deciles (identity) or by visible-token count and scene extent (support); room type and scene function are not matching variables. Because ARKitScenes classes are strongly room-correlated (stove and sink in kitchens, toilet and bathtub in bathrooms, bed in bedrooms), the +21.98-point correct-versus-donor identity effect and the +15.57-point identity-channel support effect are compatible with a room-type prior encoded in the trained predictor, with no local, object-specific completion. The class-wise donor drops in Table 7B are uniform and do not discriminate this account, and the spatial-reach diagnostic in Table 8 is inconclusive by the paper's own coverage gate (442 objects at 1.5 m, zero at 3 m). I therefore do not think the evidence currently establishes that the imputed latent contains 'context-dependent entity content' beyond a room or spatial-layout prior. Please add a same-room-type donor condition or stratify the donor effect by room type, and report a room-type probe on the imputed latent to quantify this confound.
  2. [§5.1, Table 7] The macro-average headline hides a class-level pattern that bears on the semantic-identity interpretation. Centroid-only features already exceed or nearly match SR-JEPA on several classes: sink 92.8 vs. 85.1, chair 74.6 vs. 62.4, table 41.6 vs. 42.0, and stool 40.9 vs. 31.4. Thus the +22.18-point gain over the strongest floor is concentrated in classes where simple location priors are weak, such as cabinet, bed, and bathtub. This is consistent with the predictor encoding a richer spatial or room-layout prior rather than completing the entity itself. To support the claim that the imputed latent carries entity content beyond location-based priors, please report identity accuracy conditioned on the query centroid or a residualized measure that removes a centroid-based room/location estimate.
minor comments (4)
  1. [§4.3, Appendix C] The sentence 'A preliminary coordinate-only screen made the remaining coarse families unsuitable for attribution' should state explicitly that this screen was preregistered, as claimed in Appendix C, and should list the excluded coarse families so that readers can assess the endpoint-selection risk.
  2. [Table 9 and Abstract] Since Table 9 shows the support residual is not consistently positive across checkpoints (−1.35, +2.59, +4.09), the abstract's 'unresolved 1.78-point residual' is easy to over-read; please add a qualifier such as 'pooled, with sign varying across checkpoints'.
  3. [§3.1, Eq. (2)] The stop-gradient notation 'sg(...)' is used without definition; define it in the text or a footnote.
  4. [References] Reference [15] has a broken URL line break and [17] similarly splits its URL; please fix the formatting.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the predictive pathway is trained self-supervised on ScanNet and evaluated on external ARKitScenes labels only through supervised probes; no prediction is defined in terms of its own evaluation target.

full rationale

The paper's derivation chain is self-contained against its evaluation. SR-JEPA is pretrained on ScanNet with EMA latent targets (Eq. 2) and no semantic labels, language, or 2D features. The central evaluation deletes an object, inserts a shape-free query, and reads out the frozen predictor's latent; the 43.13% identity accuracy is obtained by fitting a two-layer MLP probe on ARKitScenes Training labels and applying it to the frozen imputed latent on held-out Validation scenes. This is a standard probing protocol, not a fitted parameter renamed as a prediction: the probe is an instrument, and the claimed entity content is a property of the latent, not a quantity used to construct the latent. The composition control (predicted identity + geometry) uses a cross-fitted posterior decoded from the imputed latent; it is an empirical measurement of information content, not a definitional equivalence. The donor-context control matches token counts and bounding-box diagonals but not room type; that is a potential confounding factor for the interpretation, not a circular reduction. The paper cites prior JEPA work for architectural and objective choices, but does not rely on any self-citation for a load-bearing theorem or uniqueness claim. No equation equates a prediction with its fitted input, and the external ARKitScenes benchmark provides independent evidence. Therefore no significant circularity is present.

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

The central empirical claims rest on a label-free pretraining objective and external evaluation. The principal dependencies are hand-set query geometry, supervised probes, and annotation correctness. No new physical entities or fitted derivation constants are introduced.

free parameters (3)
  • Supervised MLP probe (identity and support heads) = 256 hidden units, GELU, dropout 0.1, AdamW lr 1e-3, up to 50 epochs
    All identity and support scores are measured by probe heads trained on labels. The endpoint is therefore a property of the latent plus the probe, not of SR-JEPA alone. The paper is transparent about this.
  • Query template (32 points, radius 0.15 m) = 32-point sphere scaled to 0.15 m
    Hand-set shape-free query geometry, identical for every object. It is not fit to labels, but the protocol's outcome depends on its scale and density.
  • Primary evaluation checkpoint and EMA schedule = epoch 95; EMA momentum 0.994 to 1.0
    The reported numbers use a chosen checkpoint fixed before endpoint scoring. Replication across checkpoints reduces, but does not eliminate, dependence on this choice.
assumptions (5)
  • domain assumption Semantic identity is faithfully measured by supervised MLP decoding from the frozen latent.
    Section 4.4 trains two-layer MLPs on frozen features; the paper's interpretation of 'entity content' assumes decodability is evidence of content.
  • domain assumption ARKitScenes, ScanNet, and Sr3D annotations and object boxes are correct and taxonomically consistent.
    Used in Sections 4.3 and Appendix C to construct deletion masks, centroid queries, and support labels.
  • domain assumption The union-hole donor construction fully removes deletion-boundary and geometry cues that distinguish correct from donor context.
    Appendix C states both arms delete the union of recipient and donor boxes and use the recipient frame, but this cannot be directly verified without code.
  • domain assumption The 32-point spherical query at the centroid carries no class-discriminative shape information.
    Section 4.1 states the template is identical for every object; the floors test location effects, but query geometry itself is not varied.
  • standard math Paired scene-cluster bootstrap resampling produces valid confidence intervals.
    Section 4.4 and Appendix C use 100,000 scene-bootstrap replicates; validity assumes scene-level clustering and correct pairing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SR-JEPA: Learning Predictive Latent State in 3D Scenes." pith.science (2026). https://pith.science/paper/N4GQVCOU

@misc{pith2026260805774,
  author       = {Pith},
  title        = {Pith review of: SR-JEPA: Learning Predictive Latent State in 3D Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N4GQVCOU}},
  note         = {Machine review of arXiv:2608.05774}
}
read the original abstract

Joint-embedding predictive architectures learn by predicting latent representations of missing observations, yet many masked JEPAs are evaluated primarily through the encoders they produce. We ask what a trained predictive pathway itself infers when an entire entity is absent from a native 3D scene. We introduce SR-JEPA, a point-native JEPA for scene-scale point clouds whose original frozen predictive pathway can be queried at a supplied location. At evaluation, every point of one object is removed before encoding and replaced by the same shape-free 32-point query at its centroid. Training uses only self-contained 3D EMA targets: no reconstruction, semantic labels, language, or lifted 2D features. On 5,953 held-out ARKitScenes objects, the imputed latent reaches 43.13% semantic-identity macro accuracy, 22.18 points above the strongest floor. Randomizing the prediction path removes 9.78 points, while substituting matched donor context removes 21.98 points. On 8,570 Sr3D support pairs, the full latent reaches 41.15 AP; identity decoded from the missing-object latent, combined with anchor identity and geometry, reaches 39.37 AP, leaving an unresolved 1.78-point residual. These results reveal a queryable, compositional 3D predictive state: the model completes context-dependent entity content, which downstream computation combines with metric geometry.

Figures

Figures reproduced from arXiv: 2608.05774 by the authors.

Figure 1
Figure 1. Query, completion, and composition expose the content of predictive 3D state. (A) The target is restored in red only for the reader; the model receives the target-deleted scene and an identical shape-free query at the supplied centroid. (B) The frozen pathway completes entity content that depends materially on predictor training and on the correct scene. The cabinet label is illustrative; effects and intervals use a… view at source ↗
Figure 2
Figure 2. SR-JEPA learns point-native 3D state by latent prediction. A context encoder sees visible occupied regions, positional mask tokens query a narrow predictor for missing latents, and an EMA target encoder processes the full scene. The marked cuts separate visible encoder state (enc), the 192-dimensional predictor embedding (emb), and completed state (pred). We retain the frozen pathway and evaluate pred directly after… view at source ↗
Figure 3
Figure 3. Completed entity state combines with geometry to support structural decisions. (A) An outcome-blind fixed-hash item from the final endpoint. Red target points are restored only for the reader; the cyan anchor remains visible. Rows give the trained vote followed by the randomized-predictor and donor-context votes. The example illustrates the downstream readout; Panels B–C test its composition. (B) Full predicted late… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Occupied-region masking remains native to the observed 3D scene. The EMA target encoder receives the complete point scene (A); cubic target regions are sampled only from occupied cells (B); and those target points are absent from the context encoder input (C). The obje…
Figure 5
Figure 5. Figure 5: Identity completion is distributed across the official ARKitScenes classes. Grouped bars show descriptive class accuracies; inset values are the claim-bearing macro effects with 95% scene-bootstrap intervals. Panels use different paired deletion geometries and should n…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 24 canonical work pages

  1. [1]

    Point2vec for self-supervised representa- tion learning on point clouds

    Karim Abou Zeid, Jonas Schult, Alexander Hermans, and Bastian Leibe. Point2vec for self-supervised representa- tion learning on point clouds. InGerman Conference on Pattern Recognition (GCPR), 2023

  2. [2]

    Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes

    Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. InECCV, 2020

  3. [3]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, 2023

  4. [4]

    V-jepa 2: Self-supervised video models enable understanding, prediction and planning

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, Sergio Arnaud, Abha Gejji, Ada Martin, Fran- cois Robert Hogan, Daniel Dugas, Piotr Bojanowski, Vasil Khalidov, Patrick Labatut, Francisco Massa, Marc Szafraniec, Kapil Krishnakumar, Yong Li, Xi...

  5. [5]

    data2vec: A general framework for self-supervised learning in speech, vision and language

    Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. InICML, pages 1298–1312, 2022

  6. [6]

    Revisiting feature prediction for learning visual representations from video (v-jepa).Transactions on Machine Learning Research (TMLR), 2024

    Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual representations from video (v-jepa).Transactions on Machine Learning Research (TMLR), 2024

  7. [7]

    Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data

    Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. InNeurIPS Datasets and Benchmarks Track, 2021

  8. [8]

    Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. InCVPR, 2017. 9 SR-JEPA: Learning Predictive Latent State in 3D ScenesA PREPRINT

Show all 33 references
  1. [9]

    Learning and leveraging world models in visual representation learning.arXiv preprint arXiv:2403.00504, 2024

    Quentin Garrido, Mahmoud Assran, Nicolas Ballas, Adrien Bardes, Laurent Najman, and Yann LeCun. Learning and leveraging world models in visual representation learning.arXiv preprint arXiv:2403.00504, 2024

  2. [10]

    Rel3d: A minimally contrastive benchmark for grounding spatial relations in 3d

    Ankit Goyal, Kaiyu Yang, Dawei Yang, and Jia Deng. Rel3d: A minimally contrastive benchmark for grounding spatial relations in 3d. InNeurIPS, 2020

  3. [11]

    Designing and interpreting probes with control tasks

    John Hewitt and Percy Liang. Designing and interpreting probes with control tasks. InEMNLP, 2019

  4. [12]

    Exploring data-efficient 3d scene understanding with contrastive scene contexts

    Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. InCVPR, 2021

  5. [13]

    3d-jepa: A joint embedding predictive architecture for 3d self-supervised representation learning.arXiv preprint arXiv:2409.15803, 2024

    Naiwen Hu, Haozhe Cheng, Yifan Xie, Shiqi Li, and Jihua Zhu. 3d-jepa: A joint embedding predictive architecture for 3d self-supervised representation learning.arXiv preprint arXiv:2409.15803, 2024

  6. [14]

    Self-supervised pre-training with masked shape prediction for 3d scene understanding

    Li Jiang, Zetong Yang, Shaoshuai Shi, Vladislav Golyanik, Dengxin Dai, and Bernt Schiele. Self-supervised pre-training with masked shape prediction for 3d scene understanding. InCVPR, pages 1168–1178, 2023

  7. [15]

    A path towards autonomous machine intelligence.OpenReview preprint, 2022

    Yann LeCun. A path towards autonomous machine intelligence.OpenReview preprint, 2022. URL https: //openreview.net/forum?id=BZ5a1r-kVsf

  8. [16]

    Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving

    Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. InIROS, 2024

  9. [17]

    Poma-3d: The point map way to 3d scene understanding

    Ye Mao, Weixun Luo, Ranran Huang, Junpeng Jing, and Krystian Mikolajczyk. Poma-3d: The point map way to 3d scene understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, pages 7282–7292, June 2026. URL https://openacce...

  10. [18]

    Locate 3d: Real-world object localization via self-supervised learning in 3d

    Paul Mcvay, Sergio Arnaud, Ada Martin, Arjun Majumdar, Krishna Murthy Jatavallabhula, Phillip Thomas, Ruslan Partsey, Daniel Dugas, Abha Gejji, Alexander Sax, Vincent-Pierre Berges, Mikael Henaff, Ayush Jain, Ang Cao, Ishita Prasad, Mrinal Kalakrishnan, Michael Rabbat, Nicolas...

  11. [19]

    V-jepa 2.1: Unlocking dense features in video self-supervised learning.arXiv preprint arXiv:2603.14482, 2026

    Lorenzo Mur-Labadia, Matthew Muckley, Amir Bar, Mido Assran, Koustuv Sinha, Mike Rabbat, Yann LeCun, Nicolas Ballas, and Adrien Bardes. V-jepa 2.1: Unlocking dense features in video self-supervised learning.arXiv preprint arXiv:2603.14482, 2026

  12. [20]

    Causal-jepa: Learning world models through object-level latent masking.arXiv preprint arXiv:2602.11389, 2026

    Heejeong Nam, Quentin Le Lidec, Lucas Maes, Yann LeCun, and Randall Balestriero. Causal-jepa: Learning world models through object-level latent masking.arXiv preprint arXiv:2602.11389, 2026

  13. [21]

    Tay, Wei Liu, Yonghong Tian, and Li Yuan

    Yatian Pang, Wenxiao Wang, Francis E.H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. InECCV, 2022

  14. [22]

    Point-jepa: A joint embedding predictive architecture for self-supervised learning on point cloud

    Ayumu Saito, Prachi Kudeshia, and Jiju Poovvancheri. Point-jepa: A joint embedding predictive architecture for self-supervised learning on point cloud. InWACV, pages 7348–7357, 2025

  15. [23]

    Learning 3d semantic scene graphs from 3d indoor reconstructions

    Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. InCVPR, 2020

  16. [24]

    Can transformers capture spatial relations between objects? In ICLR, 2024

    Chuan Wen, Dinesh Jayaraman, and Yang Gao. Can transformers capture spatial relations between objects? In ICLR, 2024

  17. [25]

    Masked scene contrast: A scalable framework for unsupervised 3d representation learning

    Xiaoyang Wu, Xin Wen, Xihui Liu, and Hengshuang Zhao. Masked scene contrast: A scalable framework for unsupervised 3d representation learning. InCVPR, 2023

  18. [26]

    Point transformer v3: Simpler, faster, stronger

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. InCVPR, 2024

  19. [27]

    Sonata: Self-supervised learning of reliable point representations

    Xiaoyang Wu, Daniel DeTone, Duncan Frost, Tianwei Shen, Chris Xie, Nan Yang, Jakob Engel, Richard New- combe, Hengshuang Zhao, and Julian Straub. Sonata: Self-supervised learning of reliable point representations. InCVPR, 2025

  20. [28]

    Qi, Leonidas Guibas, and Or Litany

    Saining Xie, Jiatao Gu, Demi Guo, Charles R. Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding. InECCV, 2020

  21. [29]

    Monocular occupancy prediction for scalable indoor scenes

    Hongxiao Yu, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Monocular occupancy prediction for scalable indoor scenes. InECCV, 2024

  22. [30]

    Point-m2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training

    Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. InNeurIPS, 2022. 10 SR-JEPA: Learning Predictive Latent State in 3D ScenesA PREPRINT

  23. [31]

    Concerto: Joint 2d-3d self-supervised learning emerges spatial representations

    Yujia Zhang, Xiaoyang Wu, Yixing Lao, Chengyao Wang, Zhuotao Tian, Naiyan Wang, and Hengshuang Zhao. Concerto: Joint 2d-3d self-supervised learning emerges spatial representations. InNeurIPS, 2025

  24. [32]

    Self-supervised jepa-based world models for lidar occupancy completion and forecasting.arXiv preprint arXiv:2602.12540, 2026

    Haoran Zhu and Anna Choromanska. Self-supervised jepa-based world models for lidar occupancy completion and forecasting.arXiv preprint arXiv:2602.12540, 2026

  25. [33]

    union-hole

    Haoran Zhu, Zhenyuan Dong, Kristi Topollai, Beiyao Sha, and Anna Ewa Choromanska. Self-supervised repre- sentation learning with joint embedding predictive architecture for automotive lidar object detection.Proceedings of the AAAI Conference on Artificial Intelligence, 40(16):...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.