Pith. sign in

REVIEW 4 major objections 3 minor 47 references

JEPADepth shows that adding a JEPA-style masked representation prediction loss to a standard photometric self-supervised depth pipeline improves depth accuracy and cross-dataset transfer, with the extra networks discarded at inference and n

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 12:34 UTC pith:PVQCQ74H

load-bearing objection A credible empirical paper showing a modest in-domain gain from an I-JEPA auxiliary loss on a DINOv3 depth pipeline, but the Cityscapes zero-shot claim is under-supported as written. the 4 major comments →

arxiv 2607.26600 v1 pith:PVQCQ74H submitted 2026-07-29 cs.CV

JEPADepth: Masked Predictive Representation Learning for Self-Supervised Monocular Depth Estimation

Ionu\c{t} Grigore , C\u{a}lin-Adrian Popa This is my paper
classification cs.CV
keywords self-supervised monocular depth estimationjoint-embedding predictive architecturemasked representation predictionDINOv3photometric reconstruction losszero-shot cross-dataset transferKITTIvision transformer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

JEPADepth asks whether a self-supervised depth network can be improved by training it to predict, in feature space, the embeddings of masked image patches rather than only reconstructing pixels through photometric warping. The paper's central claim is that adding this I-JEPA-style masked-prediction objective to a standard photometric pipeline built on a pretrained DINOv3 ViT encoder consistently improves depth accuracy and zero-shot cross-dataset transfer, with no change to the inference-time model. On KITTI, the JEPA term improves AbsRel from 0.105 to 0.101 over the same DINOv3 photometric baseline; on Make3D zero-shot it improves AbsRel from 0.284 to 0.275, and on Cityscapes it ties the best prior methods on AbsRel while leading on most other metrics. The predictor and target encoder are training-only components, so the gains come with roughly 16 percent extra training time and no deployment cost. If true, the result matters because it offers a cheap, architecture-neutral regularizer for self-supervised depth and suggests that feature-space prediction complements pixel-space geometry rather than duplicating it.

Core claim

The paper argues that photometric view-synthesis alone leaves higher-level scene structure underspecified, so a network can satisfy the pixel-warping loss while memorizing appearance statistics instead of learning geometry. JEPADepth addresses this by computing, on the patch tokens of a pretrained DINOv3 ViT, a masked prediction loss: one broad context mask and four small target masks are sampled per image; a predictor transformer infers target embeddings from visible context embeddings; an EMA-updated copy of the same encoder supplies stop-gradient targets. This loss is added to the Monodepth2-style photometric and smoothness losses with weight 1.0. The paper reports that this objective imp

What carries the argument

The load-bearing component is the JEPA-style masked representation-prediction loss (Eq. 11): a DINOv3 ViT encoder's patch tokens are split by masks into visible context and held-out target regions; a 12-layer predictor transformer maps context tokens to predicted target embeddings, and an EMA-updated target encoder provides stop-gradient targets, with an L2 match in normalized feature space. Combined with the photometric view-synthesis loss (Eq. 17), this loss forces the shared encoder to keep producing structured, semantically grounded patch features while it adapts to geometric supervision, on the hypothesis that it anchors the network to its pretrained feature manifold and prevents photom

Load-bearing premise

The moving-average target encoder, a momentum copy of the same encoder that the photometric loss is fine-tuning, keeps providing targets good enough to preserve the pretrained feature quality that makes the JEPA loss useful.

What would settle it

Run the identical JEPADepth recipe with the target encoder frozen at the original DINOv3 checkpoint, or refreshed from it, instead of EMA-updated. If KITTI AbsRel then fails to improve over the 0.105 photometric baseline, or the gain reverses, the improvement is not caused by masked prediction per se but by tracking the fine-tuning encoder; conversely, if frozen targets improve further, the EMA choice is suboptimal.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Every metric on the KITTI Eigen split improves when the JEPA objective is added to the same DINOv3-based photometric baseline (AbsRel 0.105 to 0.101, delta1 0.891 to 0.898), and the benefit appears with both FPN and DPT decoders.
  • Zero-shot transfer improves: on Make3D, AbsRel drops from 0.284 to 0.275, beating all compared methods on AbsRel and RMSE log; on Cityscapes, JEPADepth ties MonoViT on AbsRel and leads on RMSE, RMSE log, delta1, delta2, and delta3.
  • The gain is training-only: the predictor and EMA target encoder are discarded at inference, so deployment cost is identical to the photometric baseline, with about 16 percent extra training time.
  • The benefit depends on strong pretrained features: with randomly initialized DINOv3 weights the JEPA loss does not help (AbsRel 0.157 vs 0.101 pretrained), implying the objective regularizes rather than replaces the photometric signal.
  • Representation-space prediction is preferred over pixel-space reconstruction: the same loss computed in pixel space gives AbsRel 0.103 versus 0.101, consistent with the I-JEPA design principle.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The EMA target encoder is updated from the very encoder being fine-tuned by the photometric loss; the paper never compares against frozen or periodically refreshed DINOv3 targets. If EMA drift is what preserves target quality, a frozen-target variant could remove or enlarge the benefit, which would localize the mechanism.
  • Because the largest gains appear in zero-shot transfer, the JEPA term likely acts mainly as a representation regularizer, keeping features close to the pretrained manifold, rather than as a geometric constraint. A measurable prediction is that feature-space distance to DINOv3 features should correlate with the observed AbsRel improvements.
  • The same recipe should transfer to other dense prediction tasks trained by pixel-space self-supervision, such as surface-normal estimation, optical flow, or semantic segmentation, since the masked-prediction objective is agnostic to the downstream head. This is a testable extension, not claimed by the paper.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper proposes JEPADepth, which augments a standard photometric self-supervised monocular depth pipeline with an I-JEPA-style masked representation prediction loss computed on DINOv3 patch embeddings. A predictor infers target-region embeddings from context-region embeddings, and an EMA target encoder provides stop-gradient targets; both are discarded at inference, leaving a DINOv3 encoder plus a lightweight FPN depth decoder. On KITTI, the JEPA objective improves AbsRel from 0.105 to 0.101 over the matched DINOv3 photometric baseline; on zero-shot Make3D it improves AbsRel from 0.284 to 0.275; on zero-shot Cityscapes it ties MonoViT on AbsRel (0.143) and leads on most other metrics. The paper includes multiple ablations (Tables 7–11) supporting the additive gain of the JEPA objective, the importance of pretrained initialization, and the choice of representation-space prediction.

Significance. If the empirical claim holds, the contribution is practical and useful: a training-only regularizer that improves both in-domain accuracy and cross-dataset transfer with zero inference-time cost. The paper reports carefully matched ablations against the same DINOv3-based photometric baseline, tests decoder-agnostic behavior, and includes a representation-vs-pixel-space comparison. The KITTI and Make3D gains are internally consistent, and the Make3D table contains the crucial DINOv3-no-JEPA control. However, the headline differences are modest (0.105→0.101 AbsRel on KITTI; 0.284→0.275 on Make3D), no multiple-seed or error-bar information is provided, the Cityscapes table omits the DINOv3 baseline control, and the JEPA weight is selected on the same split that is used for the final KITTI numbers. These issues leave the central cross-dataset claim partially unverified rather than refuted.

major comments (4)
  1. [Table 4 / §4.5] The Cityscapes zero-shot table omits the DINOv3 w/o JEPA row that is present in the Make3D comparison (Table 5). The abstract and §4.6 attribute the cross-dataset improvement to the JEPA objective, and §4.6 explicitly uses the DINOv3 baseline to rule out the encoder alone on Make3D. Without the same control on Cityscapes, the JEPA-loss attribution is unsupported on that benchmark. If the DINOv3 no-JEPA baseline already reaches AbsRel ≤ 0.143 on Cityscapes, the central cross-dataset claim as stated in the abstract does not hold there. Add the missing baseline row.
  2. [§4.3 / Table 10] The loss weight λ_JEPA = 1.0 is selected via the sensitivity analysis in Table 10, reported on the KITTI Eigen split, and the final KITTI results in Tables 2–3 are reported on the same split. The differences involved are small (AbsRel 0.105 vs 0.101), so selecting the weight on the evaluation split risks inflating the reported gain. Please either perform the weight selection on a separate validation split or report multiple-seed results with error bars so the selection effect can be assessed.
  3. [§3.2, Eqs. (11)–(12)] The JEPA targets are outputs of an EMA copy of the same encoder that is being fine-tuned by the photometric loss, not frozen DINOv3 features. The paper hypothesizes that the JEPA loss anchors the encoder to its pretrained feature manifold, but the EMA target can drift during training. Since Table 8 shows the benefit is tightly coupled to the pretrained initialization, an ablation comparing the EMA target against frozen DINOv3 targets (or periodically reset targets) is needed to determine whether the reported gain comes from stable pretrained targets or from a self-referential regularizer. This is a concrete, checkable experiment rather than a speculative concern.
  4. [Tables 2–5] All reported numbers appear to come from single runs. The headline gains over the DINOv3 baseline are modest, and on Cityscapes the result is a tie with MonoViT on AbsRel. Without multiple-seed means and standard deviations (or a significance test), the claim of consistent additive gains and improved generalization is not robustly supported. Please report at least 3 seeds for the main comparison and for the key ablations in Tables 7 and 10.
minor comments (3)
  1. [Eq. (14)] The mean-normalized inverse depth is written as \hat{d}^*_t = \hat{d}_t / \hat{d}_t, which is identically 1. The bar over the denominator is missing; it should be divided by the mean of \hat{d}_t.
  2. [Eqs. (1)–(2)] The symbol M is overloaded: in Eq. (1) M denotes the 4×4 camera transformation matrix, while Eq. (2) uses M for the pose-network output. Additionally, t is used both as a frame index and as the translation vector t_t. Please use distinct notation.
  3. [Table 4 / §4.5] The text says JEPADepth ties MonoViT on AbsRel and is best on five of seven Cityscapes metrics, but JEPADepth has a slightly worse SqRel (1.495 vs 1.485). State this explicitly for full transparency.

Circularity Check

0 steps flagged

No significant circularity: the JEPA-depth improvement is supported by controlled ablations against a no-JEPA baseline, and no central claim reduces to its own inputs.

full rationale

The paper's central empirical claim—that adding the I-JEPA loss (Eq. 11) to a DINOv3-based photometric pipeline improves depth accuracy—is supported by controlled ablations that are not definitionally forced. Table 7 compares the same DINOv3 encoder fine-tuned with and without the JEPA objective; Table 11 repeats this comparison under a different decoder; Table 5 includes an explicit 'DINOv3 w/o JEPA' row in the Make3D zero-shot evaluation. These baselines are trained without the JEPA term, so the reported gains are not equivalent to the loss itself by construction. The target encoder is an EMA copy of the context encoder (Eq. 12), making the JEPA targets non-independent, but this is the standard I-JEPA formulation and the paper does not derive its depth results analytically from those targets; the claimed benefit is empirical. The λJEPA sensitivity sweep in Table 10 is a hyperparameter selection, not a fitted quantity renamed as a prediction, and headline results are reported on the test split described in Section 4.2. The omission of a DINOv3 w/o JEPA row in the Cityscapes table (Table 4) is an experimental-control gap that weakens the cross-dataset evidence, but it is not a circular reduction. No load-bearing self-citation appears: the authors' prior work [15] is cited only in related work, and the I-JEPA/DINOv3 components are external. The claimed improvement is therefore self-contained against external benchmarks and does not reduce to its inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim depends on standard photometric assumptions, a pretrained DINOv3 encoder, and the JEPA EMA machinery; the only tuned free parameter is λJEPA, selected on the evaluation split. No physically invented entities are introduced.

free parameters (2)
  • λ_JEPA = 1.0
    Weight of the JEPA loss in Eq. (17). Selected as best in Table 10 sensitivity analysis on the same KITTI Eigen split used for final evaluation; not derived from first principles.
  • EMA momentum schedule = 0.996 → 1.0 (linear)
    Target encoder update in Eq. (12); schedule chosen for simplicity; paper reports no significant difference with cosine, but this is a hand-set design choice.
axioms (4)
  • standard math Pinhole camera model and differentiable bilinear sampling (Eqs. 3–6).
    Used for view synthesis, the core photometric supervision; standard in the SfM depth literature.
  • domain assumption Photometric consistency (brightness constancy), static rigid scene, and auto-masking handle moving objects (Eqs. 13–16).
    Inherited from Monodepth2; the paper adopts these assumptions without re-deriving them.
  • domain assumption Pretrained DINOv3 features provide meaningful JEPA prediction targets (Table 8).
    Ablation shows the JEPA gain nearly disappears with random initialization, so the whole method depends on this external pretrained representation.
  • domain assumption EMA target encoder with stop-gradient prevents collapse and preserves target quality during joint photometric+JEPA fine-tuning (Eq. 12).
    Critical design assumption; the paper does not ablate frozen vs EMA targets, so target drift during fine-tuning is unexamined.

pith-pipeline@v1.3.0-daily-deepseek · 15398 in / 15680 out tokens · 157663 ms · 2026-08-01T12:34:42.319101+00:00 · methodology

0 comments
read the original abstract

Self-supervised monocular depth estimation typically relies on photometric reconstruction losses that couple depth, pose, and appearance assumptions. In this paper, we propose JEPADepth, a self-supervised monocular depth framework that incorporates a complementary training objective inspired by Image Joint-Embedding Predictive Architectures (I-JEPA) for self-supervised depth learning. Our method augments a standard photometric pipeline with a masked prediction loss computed in the representation space of a pretrained DINOv3 Vision Transformer encoder. A predictor infers target-region embeddings from visible context-region embeddings under structured masking, and is discarded along with the target encoder at inference time, adding no deployment cost. On KITTI, adding the JEPA objective consistently improves performance over the same DINOv3-based photometric baseline, without changing the inference-time architecture. Compared to prior monocular self-supervised methods, JEPADepth is competitive with state-of-the-art transformer-based approaches and outperforms strong CNN-based baselines on the standard benchmark. In zero-shot transfer (trained on KITTI and evaluated without fine-tuning), JEPADepth achieves the best or near-best performance among the compared methods on both Make3D and Cityscapes across multiple metrics.

Figures

Figures reproduced from arXiv: 2607.26600 by C\u{a}lin-Adrian Popa, Ionu\c{t} Grigore.

Figure 1
Figure 1. Figure 1: Overview of JEPADepth. A DINOv3 ViT encoder extracts patch tokens from the target image. A context mask selects visible tokens (blue) that are fed to a predictor, which generates embeddings for multiple masked target regions (red). In parallel, an EMA-updated target encoder produces target-region embeddings used as stop-gradient learning targets. The predicted and target embeddings are matched with a JEPA … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Image-JEPA training framework [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Context-target masking strategy. For each input image, we randomly sam￾ple four target regions with scales drawn from the interval (0.15, 0.2) and aspect ratios from (0.75, 1.5). Subsequently, we sample a context region with scale in (0.85, 1.0) and discard any overlapping target regions. This design encourages the target regions to capture semantically meaningful content, while the context region remains … view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of model efficiency and depth accuracy on KITTI using [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: shows qualitative depth predictions on the KITTI Eigen split. The JEPA objective yields sharper object boundaries and more coherent scene struc￾ture than the photometric-only DINOv3 baseline [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative depth estimation results on Make3D (zero-shot). [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 10 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N.: Self-supervised learning from images with a joint-embedding predic- tive architecture. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15619–15629 (2023)

    2023

  2. [2]

    arXiv preprint arXiv:2506.09985 (2025)

    Assran, M., Bardes, A., Fan, D., Garrido, Q., Howes, R., Muckley, M., Rizvi, A., Roberts, C., Sinha, K., Zholus, A., et al.: V-jepa 2: Self-supervised video models enable understanding, prediction and planning. arXiv preprint arXiv:2506.09985 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops

    Casser, V., Pirk, S., Mahjourian, R., Angelova, A.: Unsupervised monocular depth and ego-motion learning with structure and semantics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0–0 (2019)

    2019

  4. [4]

    In: 2023 IEEE International Conference on Robotics and Automation (ICRA)

    Chawla, H., Jeeveswaran, K., Arani, E., Zonooz, B.: Image masking for robust self- supervised monocular depth estimation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 10054–10060. IEEE (2023)

    2023

  5. [5]

    arXiv preprint arXiv:2502.14573 (2025)

    Choi, W., Hwang, K., Peng, W., Choi, M., Im, S.: Self-supervised monocular depth estimation robust to reflective surface leveraged by triplet mining. arXiv preprint arXiv:2502.14573 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213–3223 (2016)

    2016

  7. [7]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  8. [8]

    In: Proceedings of the IEEE in- ternational conference on computer vision

    Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE in- ternational conference on computer vision. pp. 2650–2658 (2015)

    2015

  9. [9]

    Advances in neural information processing systems27 (2014)

    Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems27 (2014)

    2014

  10. [10]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Gao, X., Wu, X., Ning, Z., Yang, R., Zheng, Z., Yang, J., Liu, W.: Adadepth: Ex- ploiting inherent scene information for self-supervised depth estimation in dynamic scenes. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 4185–4193 (2026)

    2026

  11. [11]

    The international journal of robotics research32(11), 1231–1237 (2013)

    Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The international journal of robotics research32(11), 1231–1237 (2013)

    2013

  12. [12]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth es- timation with left-right consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 270–279 (2017)

    2017

  13. [13]

    In: Proceedings of the IEEE/CVF in- ternational conference on computer vision

    Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self- supervised monocular depth estimation. In: Proceedings of the IEEE/CVF in- ternational conference on computer vision. pp. 3828–3838 (2019)

    2019

  14. [14]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Gordon, A., Li, H., Jonschkowski, R., Angelova, A.: Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8977–8986 (2019) JEPADepth 21

    2019

  15. [15]

    arXiv preprint arXiv:2406.04532 (2024)

    Grigore, I., Popa, C.A.: Mambadepth: Enhancing long-range dependency for self-supervised fine-structured monocular depth estimation. arXiv preprint arXiv:2406.04532 (2024)

    Pith/arXiv arXiv 2024

  16. [16]

    arXiv preprint arXiv:1905.026935(1) (2019)

    Guizilini, V., Ambrus, R., Pillai, S., Gaidon, A.: Packnet-sfm: 3d packing for self-supervised monocular depth estimation. arXiv preprint arXiv:1905.026935(1) (2019)

    Pith/arXiv arXiv 1905

  17. [17]

    Han, W., Yin, J., Shen, J.: Self-supervised monocular depth estimation by direction-awarecumulativeconvolutionnetwork.In:ProceedingsoftheIEEE/CVF International Conference on Computer Vision. pp. 8613–8623 (2023)

    2023

  18. [18]

    IEEE Transactions on pattern analysis and machine intelligence30(2), 328– 341 (2007)

    Hirschmuller, H.: Stereo processing by semiglobal matching and mutual informa- tion. IEEE Transactions on pattern analysis and machine intelligence30(2), 328– 341 (2007)

    2007

  19. [19]

    Advances in neural information processing systems28(2015)

    Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. Advances in neural information processing systems28(2015)

    2015

  20. [20]

    In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition

    Johnston, A., Carneiro, G.: Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition. pp. 4756–4765 (2020)

    2020

  21. [21]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  22. [22]

    arXiv preprint arXiv:2109.06163 (2021)

    Li, Z., Drenkow, N., Ding, H., Ding, A.S., Lu, A., Creighton, F.X., Taylor, R.H., Unberath, M.: On the sins of image synthesis loss for self-supervised depth esti- mation. arXiv preprint arXiv:2109.06163 (2021)

    Pith/arXiv arXiv 2021

  23. [23]

    In: Proceedings of the AAAI conference on artificial intelligence

    Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hr- depth: High resolution self-supervised monocular depth estimation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 2294–2301 (2021)

    2021

  24. [24]

    Advances in neural information processing sys- tems32(2019)

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Advances in neural information processing sys- tems32(2019)

    2019

  25. [25]

    In: 2018 international conference on 3D vision (3DV)

    Pilzer, A., Xu, D., Puscas, M., Ricci, E., Sebe, N.: Unsupervised adversarial depth estimation using cycled generative networks. In: 2018 international conference on 3D vision (3DV). pp. 587–595. IEEE (2018)

    2018

  26. [26]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179–12188 (2021)

    2021

  27. [27]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Com- petitivecollaboration:Jointunsupervisedlearningofdepth,cameramotion,optical flow and motion segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12240–12249 (2019)

    2019

  28. [28]

    Saito, A., Kudeshia, P., Poovvancheri, J.: Point-jepa: A joint embedding predictive architectureforself-supervisedlearningonpointcloud.In:2025IEEE/CVFWinter Conference on Applications of Computer Vision (WACV). pp. 7348–7357. IEEE (2025)

    2025

  29. [29]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Saunders, K., Vogiatzis, G., Manso, L.J.: Self-supervised monocular depth estima- tion: Let’s talk about the weather. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8907–8917 (2023)

    2023

  30. [30]

    In: 2007 IEEE 11th international conference on computer vision

    Saxena, A., Sun, M., Ng, A.Y.: Learning 3-d scene structure from a single still image. In: 2007 IEEE 11th international conference on computer vision. pp. 1–8. IEEE (2007) 22 I. Grigore and C.-A. Popa

    2007

  31. [31]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Shi, X., Dikov, G., Reitmayr, G., Kim, T.K., Ghafoorian, M.: 3d distillation: Improving self-supervised monocular depth estimation on reflective surfaces. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9133–9143 (2023)

    2023

  32. [32]

    In: European Conference on Computer Vision

    Shu, C., Yu, K., Duan, Z., Yang, K.: Feature-metric loss for self-supervised learning of depth and egomotion. In: European Conference on Computer Vision. pp. 572–

  33. [33]

    arXiv preprint arXiv:2508.10104 (2025)

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khali- dov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. arXiv preprint arXiv:2508.10104 (2025)

    Pith/arXiv arXiv 2025

  34. [34]

    IEEE Transactions on Neural Networks and Learning Systems 34(12), 9604–9624 (2022)

    Tang, Y., Zhao, C., Wang, J., Zhang, C., Sun, Q., Zheng, W.X., Du, W., Qian, F., Kurths, J.: Perception and navigation in autonomous systems in the era of learning: A survey. IEEE Transactions on Neural Networks and Learning Systems 34(12), 9604–9624 (2022)

    2022

  35. [35]

    Terver, B., Yang, T.Y., Ponce, J., Bardes, A., LeCun, Y.: What drives success in physical planning with joint-embedding predictive world models? arXiv preprint arXiv:2512.24497 (2025)

    Pith/arXiv arXiv 2025

  36. [36]

    In: 2017 international conference on 3D Vision (3DV)

    Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 2017 international conference on 3D Vision (3DV). pp. 11–20. IEEE (2017)

    2017

  37. [37]

    In: Conference on Robot Learning

    Vankadari, M., Golodetz, S., Garg, S., Shin, S., Markham, A., Trigoni, N.: When the sun goes down: Repairing photometric losses for all-day depth estimation. In: Conference on Robot Learning. pp. 1992–2003. PMLR (2023)

    1992

  38. [38]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2022–2030 (2018)

    2022

  39. [39]

    In: Proceedings of the IEEE/CVF international confer- ence on computer vision

    Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: Proceedings of the IEEE/CVF international confer- ence on computer vision. pp. 2162–2171 (2019)

    2019

  40. [40]

    In: 2021 International Conference on 3D vision (3DV)

    Yan, J., Zhao, H., Bu, P., Jin, Y.: Channel-wise attention-based network for self- supervised monocular depth estimation. In: 2021 International Conference on 3D vision (3DV). pp. 464–473. IEEE (2021)

    2021

  41. [41]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhan, H., Garg, R., Weerasekera, C.S., Li, K., Agarwal, H., Reid, I.: Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 340–349 (2018)

    2018

  42. [42]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18537–18546 (2023)

    2023

  43. [43]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhang, W., Liu, H., Li, B., He, J., Qi, Z., Wang, Y., Zhao, S., Yu, X., Zeng, W., Jin, X.: Hybrid-grained feature aggregation with coarse-to-fine language guidance for self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6678–6692 (2025)

    2025

  44. [44]

    Science China Technological Sciences63(9), 1612–1627 (2020)

    Zhao, C., Sun, Q., Zhang, C., Tang, Y., Qian, F.: Monocular depth estimation based on deep learning: An overview. Science China Technological Sciences63(9), 1612–1627 (2020)

    2020

  45. [45]

    In: 2022 international conference on 3D vision (3DV)

    Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 2022 international conference on 3D vision (3DV). pp. 668–678. IEEE (2022) JEPADepth 23

    2022

  46. [46]

    arXiv preprint arXiv:2110.09482 (2021)

    Zhou, H., Greenwood, D., Taylor, S.: Self-supervised monocular depth estimation with internal feature fusion. arXiv preprint arXiv:2110.09482 (2021)

    Pith/arXiv arXiv 2021

  47. [47]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1851–1858 (2017)

    2017