CoChat
Sign Up
Cited by 2,561Advanced Vision and ImagingRobotics and Sensor-Based LocalizationAdvanced Neural Network Applications

Digging Into Self-Supervised Monocular Depth Estimation

Open access
Godard, Clement, Aodha, Oisin Mac, Firman, Michael, Brostow, Gabriel·2019·2019 IEEE/CVF International Conference on Computer Vision (ICCV)
DOI 10.1109/iccv.2019.00393
Open this paper in CoChat — free accountOpen PDF

Abstract

Per-pixel ground-truth depth data is challenging to acquire at scale. To overcome this limitation, self-supervised learning has emerged as a promising alternative for training models to perform monocular depth estimation. In this paper, we propose a set of improvements, which together result in both quantitatively and qualitatively improved depth maps compared to competing self-supervised methods. Research on self-supervised monocular training usually explores increasingly complex architectures, loss functions, and image formation models, all of which have recently helped to close the gap with fully-supervised methods. We show that a surprisingly simple model, and associated design choices, lead to superior predictions. In particular, we propose (i) a minimum reprojection loss, designed to robustly handle occlusions, (ii) a full-resolution multi-scale sampling method that reduces visual artifacts, and (iii) an auto-masking loss to ignore training pixels that violate camera motion assumptions. We demonstrate the effectiveness of each component in isolation, and show high quality, state-of-the-art results on the KITTI benchmark.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper introduces Monodepth2, a self-supervised monocular depth estimation model with three simple changes: per-pixel minimum reprojection loss, auto-masking of stationary pixels, and full-resolution multi-scale sampling. It reports state-of-the-art self-supervised depth results on KITTI and shows each component improves over a baseline.

Key findings

  • On the KITTI Eigen split, Monodepth2 with monocular training reaches Abs Rel 0.115, Sq Rel 0.903, RMSE 4.863, RMSE log 0.193, and δ<1.25 0.877; with monocular+stereo training it reaches Abs Rel 0.106, RMSE 4.750, and δ<1.25 0.874.
  • Ablation shows the baseline without the three contributions performs worst (Abs Rel 0.140, RMSE 5.512), while the full model improves to Abs Rel 0.115, RMSE 4.863.
  • Even without ImageNet pretraining, the model remains competitive: monocular Abs Rel 0.132, RMSE 5.142; stereo Abs Rel 0.130, RMSE 5.485; monocular+stereo Abs Rel 0.127, RMSE 5.266.
  • On Make3D, Monodepth2 with monocular training reports Abs Rel 0.322, Sq Rel 3.589, RMSE 7.417, RMSE log 0.163, outperforming methods that do not use depth supervision.

Methods

  • Self-supervised training by minimizing photometric reprojection error between a target image and source views, using predicted depth and camera pose; monocular training uses temporally adjacent frames, stereo uses the opposite stereo view, and mixed training uses both.
  • Depth network is a U-Net encoder-decoder with a ResNet18 encoder (11M parameters); pose network is a separate ResNet18 modified to accept two frames/six channels and predict a 6-DoF relative pose.
  • Trained in PyTorch for 20 epochs with Adam, batch size 12, input/output resolution 640×192, learning rate 10^-4 for 15 epochs then 10^-5, smoothness weight λ=0.001; KITTI Eigen split uses 39,810 monocular triplets for training and 4,424 for validation.

Limitations

  • The paper states the model breaks when scenes contain objects that violate the Lambertian assumptions of the appearance loss, and shows failure cases on distorted, reflective, and color-saturated regions, ambiguous boundaries, and intricate shapes.
  • The paper cautions that Make3D ground truth depth and input images are not well aligned, causing potential evaluation issues.
  • The paper notes monocular training has no guarantee of metric scale; per-image median scaling can hide unstable scale estimation, and the authors propose single-scale evaluation as fairer but still using ground-truth depth.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/iccv.2019.00393

Cites

80 of 80 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 29 cited sources checked · 0 supported · 0 flagged

  • Findingcitation

    Reflection padding significantly reduces border artifacts found in existing approaches.

    “We found that this significantly reduces the border artifacts found in existing approaches, e.g.[15].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [15] source

      Not checked in the free run.

  • Findingown data

    The pose network performs on par with competing methods on KITTI odometry.

    “While our focus is better depth estimation, our pose network performs on par with competing methods.”
    • Also stated in: § A. Odometry Evaluation · hedged
      “As we can see in Table 4, our frame-to-frame poses come close to the accuracy of methods trained on blocks of five frames at a time.”
    Quote in the paperNot a substantial claim
  • Findingnone given

    With careful choices regarding appearance losses and image resolution, monocular training can reach the performance of stereo training.

    “In this work, we show that with careful choices regarding appearance losses and image resolution, we can reach the performance of stereo training using only monocular training.”
    Quote in the paperSubstantial claim
  • Findingown data

    Replacing auto-masking with a re-implementation of the predictive mask from [76] is worse than using no masking at all.

    “We find that this ablated model is worse than using no masking at all, while our auto-masking improves results in all cases.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [76] source

      Not checked in the free run.

  • Findingown data

    The method outperforms recent methods that explicitly compute optical flow and motion masks.

    “We also outperform recent methods ( [38,51]) that explicitly compute optical flow as well as motion masks.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [38] source

      Not checked in the free run.

    • cites [51] source

      Not checked in the free run.

  • May combine two claims. 17
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 8

  • May combine two claims.

    Findingown data

    The method achieves high accuracy in the stereo-only case despite using lower resolution and substantially less training time than [47].

    “We achieve high accuracy despite using a lower resolution than [47]'s 1024 × 384, with substantially less training time (20 vs. 200 epochs) and no use of postprocessing.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [47] source

      Not checked in the free run.

  • May combine two claims.

    Findingnone given

    The proposed improvements significantly increase predicted depth quality without adding additional model components that also require training.

    “we propose several improvements that significantly increase predicted depth quality, without adding additional model components that also require training”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Existing monocular methods produce lower quality depths than the best fully-supervised models.

  • May combine two claims.

    Findingown data

    All components together give the highest quality model, and pretraining helps.

    “We confirm the finding of the main paper, that all our components together gives the highest quality model, and that pretraining helps.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    The method shows superior performance on the KITTI Depth Prediction Benchmark using improved ground truth depths.

    “Like [1], we use these instead of the reprojected LIDAR scans to compare our method against several existing baseline algorithms, still showing superior performance.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] source

      Not checked in the free run.

  • May combine two claims.

    Findingnone given

    The model can be trained with monocular video data, stereo data, or mixed monocular and stereo data.

    “We showed how together they give a simple and efficient model for depth estimation, which can be trained with monocular video data, stereo data, or mixed monocular and stereo data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The minimum reprojection loss significantly reduces artifacts at image borders, improves the sharpness of occlusion boundaries, and leads to better accuracy.

    “Using our minimum reprojection loss significantly reduces artifacts at image borders, improves the sharpness of occlusion boundaries, and leads to better accuracy (see Table 2).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The shared pose encoder resulted in fewer parameters to optimize during training, but also results in a decrease in depth prediction accuracy.

    “This resulted in fewer parameters to optimize during training, but also results in a decrease in depth prediction accuracy, see Table 12.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Our models generate higher quality outputs and do not produce holes in the depth maps or border artifacts that can be seen in many existing baselines.

    “We can see that our models generate higher quality outputs and do not produce 'holes' in the depth maps or border artifacts that can be seen in many existing baselines e.g.[76,51,15,73].”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [76] source

      Not checked in the free run.

    • cites [51] source

      Not checked in the free run.

    • cites [15] source

      Not checked in the free run.

    • cites [73] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Using a feature reconstruction loss takes significantly longer to train but results in qualitatively better depth maps on the Wander dataset.

    “This takes significantly longer to train, but results in qualitatively better depth maps on this dataset.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Our method outperforms previously published self-supervised monocular methods on single-scale evaluation, especially in the near range depth values δ < 1.25, and is more stable overall.

    “As can be seen in Table 8, our method outperforms previously published self-supervised monocular methods, especially in the near range depth values i.e. δ < 1.25, and is more stable overall.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Our separate encoder is superior for both pretrained and non-pretrained variants, whether we use two or three frames as input.

    “Overall, our separate encoder is superior for both pretrained and non-pretrained variants, whether we use two or three frames as input.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    At the time of writing, there were no published self-supervised approaches among the submissions on the KITTI leaderboard.

    “At the time of writing, there were no published self-supervised approaches among the submissions on the leaderboard.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The method simplifies many components found in existing top performing models.

    “and simplify many components found in the existing top performing models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The auto-masking lets the network ignore objects moving at the same velocity as the camera and ignore whole frames when the camera stops moving.

    “This has the effect of letting the network ignore objects which move at the same velocity as the camera, and even to ignore whole frames in monocular videos when the camera stops moving.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Self-supervised monocular training operates under the assumptions of a moving camera and a static scene.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The learning rate was chosen using a dedicated validation set of 10% of the data.

    “This was chosen using a dedicated validation set of 10% of the data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    As a baseline we compare against the pose network used by [62], which builds upon [76] with an additional scaling of the translation by the mean of the inverse depth.

    “As a baseline we compare against the pose network used by [62], which builds upon [76] with an additional scaling of the translation by the mean of the inverse depth.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [62] source

      Not checked in the free run.

    • cites [76] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Post-processing, introduced by [15], is a technique to improve test time results on stereo-trained monocular depth estimation methods by running each test image through the network twice, once unflipped and then flipped.

    “estimation methods by running each test image through the network twice, once unflipped and then flipped.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [15] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Post-processing has been shown to bring significant gains in accuracy for stereo results, at the expense of requiring two forward-passes through the network at test time.

    “This has been shown to bring significant gains in accuracy for stereo results, at the expense of requiring two forward-passes through the network at test time [15,50].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [15] source

      Not checked in the free run.

    • cites [50]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    The Wander dataset is quite different from the car mounted videos of KITTI as it only features a monocular hand-held camera in a non-European environment.

    “This dataset is quite different from the car mounted videos of KITTI as it only features a monocular hand-held camera in a non-European environment.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Color augmentations were only applied to the images fed to the networks, not to those used to compute the photometric loss.

    “Importantly, the color augmentations are only applied to the images which are fed to the networks, not to those used to compute L p .”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    [59] introduced a set of high quality depth maps for the KITTI dataset, making use of 5 consecutive frames and handling moving objects using the stereo pair.

    “[59] introduced a set of high quality depth maps for the KITTI dataset, making use of 5 consecutive frames and handling moving objects using the stereo pair.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [59]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The evaluation method introduced by Eigen [8] for KITTI uses reprojected LIDAR points but does not handle occlusions, moving objects, or the fact that the car is moving.

    “The evaluation method introduced by Eigen [8] for KITTI uses the reprojected LI-DAR points but does not handle occlusions, moving objects, or the fact that the car is moving.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Per-image depth scaling hides unstable scale estimation in both depth and pose estimation and presents a best-case scenario for the monocular training case.

    “This per-image depth scaling hides unstable scale estimation in both depth and pose estimation and presents a best-case scenario for the monocular training case.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Monocular self-supervised depth estimation methods have no guarantee of producing metric-scale depth.
23 of 23

  • The paper relies on this without support.

    Per-pixel ground-truth depth data is challenging to acquire at scale.

    “Per-pixel ground-truth depth data is challenging to acquire at scale.”

  • The authors state this as an assumption.

    Self-supervised training relies on assumptions of brightness constancy and Lambertian material properties of object surfaces between frames.

    “Self-supervised training typically relies on making assumptions about the appearance (i.e.brightness constancy) and material properties (e.g.Lambertian) of object surfaces between frames.”

  • The authors state this as an assumption.

    Self-supervised monocular training operates under the assumptions of a moving camera and a static scene.

    “Self-supervised monocular training often operates under the assumptions of a moving camera and a static scene.”

  • The authors state this as an assumption.

    The pre-computed intrinsics K of all views are identical.

    “For simplicity of notation we assume the pre-computed intrinsics K of all the views are identical, though they can be different.”

  • The paper relies on this without support.

    Existing monocular methods produce lower quality depths than the best fully-supervised models.

    “Existing monocular methods produce lower quality depths than the best fully-supervised models.”

  • The paper relies on this without support.

    Monocular self-supervised depth estimation methods have no guarantee of producing metric-scale depth.

    “Our monocular trained approach, like all self-supervised baselines, has no guarantee of producing results with a metric scale.”
6 of 6

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Deep-learning-based visual odometry accuracy is inferior to traditional methods because it ignores geometric information.

    “More surprisingly, they show that the welltrained networks enable scale-consistent predictions over long videos, while the accuracy is still inferior to traditional methods because of ignoring geometric information.”

    From DF-VO: What Should Be Learnt for Visual Odometry? (2021) · cites this paper as Godard et al. 2019

    The odometry results in B. Network Details (page unavailable) show Monodepth2 with higher trajectory error than the traditional ORB-SLAM baseline on both KITTI test sequences. This supports a narrower accuracy comparison, but the paper does not attribute that difference to ignoring geometric information or establish it for deep-learning visual odometry generally. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

  • Well-trained deep networks enable scale-consistent predictions over long videos.

    “More surprisingly, they show that the welltrained networks enable scale-consistent predictions over long videos, while the accuracy is still inferior to traditional methods because of ignoring geometric information.”

    From DF-VO: What Should Be Learnt for Visual Odometry? (2021) · cites this paper as Godard et al. 2019

    Section D.2, “Single-Scale Evaluation” (page unavailable), reports that the authors’ monocular depth predictions have more stable scales than previously published self-supervised methods. It does not establish scale consistency over long videos: the odometry evaluation in Section A measures overlapping five-frame snippets, not long-video predictions. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

  • Feature-metric losses prevent the training objective from getting stuck at multiple local minima for homogeneous areas.

    “The main goal of these losses is to prevent the training objective from getting stuck at multiple local minima for homogeneous areas as fisheye images have considerably larger homogeneous areas than their rectilinear counterpart.”

    From OmniDet: Surround View Cameras Based Multi-Task Visual Perception Network for Autonomous Driving (2021) · cites this paper as [11]

    The Multi-scale Estimation section says multi-scale depth prediction and image reconstruction are used to keep training from getting stuck in local minima, and discusses ambiguity in large low-texture regions. It does not attribute that effect to feature-metric losses or compare homogeneous areas in fisheye and rectilinear images. Page numbers are unavailable in our copy. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • The main supervision signal in the self-supervised framework is photometric consistency across multiple views.

    “Instead of using ground truth supervisions, the main supervision signal in the selfsupervised framework is photometric consistency across multiple-views.”

    From DF-VO: What Should Be Learnt for Visual Odometry? (2021) · cites this paper as Godard et al. 2019

  • Using the minimum photometric error between the reference pixel and the synthesized pixel, instead of averaging over multiple views, overcomes out-of-view pixels and occlusions.

    “As proposed in (Godard et al. 2019), instead of averaging the photometric errors between the reference pixel and the synthesized pixels from multiple views, (Godard et al. 2019) only counts the photometric error between the reference pixel and the synthesized pixel with the minimum error. The rationale is to overcome the issues related to out-of-view pixels and occlusions.”

    From DF-VO: What Should Be Learnt for Visual Odometry? (2021) · cites this paper as Godard et al. 2019

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • SuperDepth input height is reported inconsistently

    Minor

    Table 1 gives SuperDepth's resolution as 1024 × 382, whereas the text describes the same cited model's resolution as 1024 × 384.

  • High-resolution stereo results differ between tables

    Minor

    Table 1 reports Abs Rel 0.107, Sq Rel 0.849, and RMSE 4.764 for the high-resolution stereo model, while Table 11 reports 0.105, 0.822, and 4.692 for the same model and resolution. The Table 11 values instead match the post-processed results in Table 10, although its row is not labeled as post-processed.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 9

  1. zbontar, 2016, Stereo matching by training a convolutional neural network to compare image patches, JMLR
  2. vijayanarasimhan, 2017, SfMNet: Learning of structure and motion from video, ArXiv
  3. 2017, KITTI Single Depth Evaluation Server
  4. xie, 2016, Deep3D: Fully automatic 2D-to-3D video conversion with deep convolutional neural networks, ECCV
  5. ilg, 2017, FlowNet2: Evolution of optical flow estimation with deep networks, CVPR
  6. jaderberg, 2015, Spatial transformer networks, NeurIPS
  7. simonyan, 2015, Very deep convolutional networks for large-scale image recognition, ICLRE
  8. ronneberger, 2015, UNet: Convolutional networks for biomedical image segmentation, MICCAI
  9. chen, 2016, Single-image depth perception in the wild, NeurIPS

Cited by

Sorted by citations

Related

Authors
Godard, C., Aodha, O. M., Firman, M., & Brostow, G.
Affiliations
  • California Institute of Technology (US)
  • University College London (GB)
ORCID
Venue
2019 IEEE/CVF International Conference on Computer Vision (ICCV)
Publisher
IEEE
Published
2019-10
Pages
3827-3837
DOI
10.1109/iccv.2019.00393
OpenAlex
W2806446538
Semantic Scholar
589cfcb2f995c94b0a98c902cc1f5e0f27cbd927
Cited by
2,561
References
80
Citations by year
2026: 212 · 2025: 377 · 2024: 488 · 2023: 460 · 2022: 383 · 2021: 386 · 2020: 196 · 2019: 52 · 2018: 7
FWCI
82.95
Type
conference paper
Language
en
Open access
Licence
Locations
  • IEEE/CVF International Conference on Computer Vision (ICCV) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • CaltechAUTHORS (California Institute of Technology) (submittedVersion)
  • UCL Discovery (University College London) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Vision and Imaging
  • Physical Sciences › Engineering › Aerospace Engineering › Robotics and Sensor-Based Localization
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
Keywords
monocular depth estimation, self-supervised learning, occlusion handling, multi-scale sampling, visual artifacts, KITTI dataset

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review