Back

Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue

Ravi GargVijay Kumar BGGustavo CarneiroIan Reid

article2016ECCV1,664 citations

Proposes an unsupervised framework for single-view depth estimation that trains a convolutional neural network using image reconstruction loss across stereo pairs, matching the accuracy of fully supervised models on KITTI without requiring ground-truth depth data.

Listen

Estimating 3D depth from a single 2D image is critical for applications such as autonomous driving, robotics, and augmented reality. Conventional deep learning methods for single-view depth estimation require massive datasets paired with ground-truth 3D depth measurements. Acquiring this ground truth requires expensive specialized hardware like laser scanners, precise sensor calibration, and extensive manual effort, and models trained on such data often fail to adapt when deployed in new operating environments.

The article demonstrates an unsupervised deep learning framework capable of predicting accurate depth maps from a single image without requiring ground-truth depth data or pre-trained models. The primary objective is to show that a neural network can learn single-view 3D depth estimation end-to-end using solely unannotated stereo image pairs captured with a standard binocular camera rig.

The researchers formulated the training pipeline analogously to an autoencoder that incorporates geometric image warping. The neural network acts as an encoder, taking a single left image as input and predicting its depth map. Instead of training a separate decoder network, the system uses the predicted depth and known camera geometry to mathematically warp the corresponding right stereo image to reconstruct the original left image. The network minimizes the photometric difference between the reconstructed image and the actual left image alongside a smoothness constraint to resolve ambiguous regions. The method was evaluated on outdoor driving scenes from the public KITTI benchmark, using 22,600 stereo pairs for training and 697 images for testing across standard quantitative depth error metrics.

The findings show that the proposed unsupervised network achieves single-view depth estimation accuracy that matches or exceeds existing state-of-the-art supervised systems. First, when fine-tuned with data augmentation, the model achieved a root mean squared error of 5.104 meters and a squared relative error of 1.080, significantly outperforming supervised baselines that were trained directly on ground-truth laser data. Second, the unsupervised framework captured fine geometric details, such as traffic lights and pedestrians, that were completely missed by supervised models. Third, the framework outperformed supervised networks trained on proxy depths generated by traditional stereo algorithms, proving that direct photometric image reconstruction prevents the network from learning the systematic stereo matching errors and data voids common in classic stereo pipelines.

These results demonstrate that organizations can train high-performing computer vision depth estimation systems without costly 3D sensors or human data labeling. Eliminating the dependency on ground-truth annotations significantly reduces data collection costs, accelerates deployment timelines, and facilitates continuous, in-situ model adaptation in production environments.

Organizations developing perception pipelines should consider adopting unsupervised stereo-based training frameworks to scale their training workflows and lower data acquisition expenses. Before full operational deployment, engineering teams should conduct pilot evaluations to test the framework on continuous camera streams and monocular navigation feeds, which will establish system performance across varying lighting and weather conditions. Future technical initiatives should focus on testing higher-capacity network architectures and incorporating edge-preserving smoothness formulations to further refine object boundaries.

Key limitations include the use of a basic smoothness regularizer that can cause slight blurring at distant object boundaries, as well as testing that was confined to rectified automotive stereo camera setups clamped to an evaluation range of 50 meters. Confidence in the reported results is high for outdoor street-level navigation, but performance should be re-evaluated when applying the method to indoor scenes or uncalibrated single-camera setups.

University of Adelaideautoencodersconvolutional encodersconvolutional neural networkssingle view depth estimationunsupervised CNNdepth prediction network
arXiv: 1603.04992Ravi-Garg/Unsupervised_Depth_Estimation
Cover for Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue

Abstract

A significant weakness of most current deep Convolutional Neural Networks is the need to train them using vast amounts of manu- ally labelled data. In this work we propose a unsupervised framework to learn a deep convolutional neural network for single view depth predic- tion, without requiring a pre-training stage or annotated ground truth depths. We achieve this by training the network in a manner analogous to an autoencoder. At training time we consider a pair of images, source and target, with small, known camera motion between the two such as a stereo pair. We train the convolutional encoder for the task of predicting the depth map for the source image. To do so, we explicitly generate an inverse warp of the target image using the predicted depth and known inter-view displacement, to reconstruct the source image; the photomet- ric error in the reconstruction is the reconstruction loss for the encoder. The acquisition of this training data is considerably simpler than for equivalent systems, requiring no manual annotation, nor calibration of depth sensor to camera. We show that our network trained on less than half of the KITTI dataset (without any further augmentation) gives com- parable performance to that of the state of art supervised methods for single view depth estimation.

Topics

autoencodersconvolutional encodersconvolutional neural networksdepth estimationgeometryreconstructionstereo-imageviewpointsource imagesparallel performance

Keywords

single view depth estimationunsupervised CNNdepth prediction networkdepth mapstereo matchingoptical flowmonocular imagesinverse depthunsupervised training

Table of Contents

  • 1 Introduction
  • 2 Approach
  • 2.1 Autoencoder loss
  • 3 Coarse-to-fine training with skip architecture
  • 4 Network Architecture
  • 5 Experiments
  • 5.1 Implementation Details
  • 5.2 Effect of upsampling
  • 5.3 Fine tuning with augmentation
  • 5.4 Comparison with State-of-the-art Methods on KITTI Dataset
  • 5.5 Comparisons with Baseline Supervised Networks and Stereo
  • 6 Related work
  • 7 Conclusions
  • References

Knowls

  1. Knowl 1 — Stereopsis-Based Autoencoder Formulation for Unsupervised Single-View Depth Estimation

    model/method

    The single-view depth estimation task is formulated as an unsupervised autoencoder trained using rectified stereo image pairs {I1,I2}\{I_1, I_2\} with known baseline separation BB and camera focal length ff.

    The encoder is a convolutional neural network that accepts only the single source image I1I_1 as input and predicts a per-pixel dense disparity map D(x)D(x) (or equivalently per-pixel depth d(x)=fB/D(x)d(x) = fB / D(x)). The decoder is a deterministic, non-learned geometric backward-warping operation that reconstructs the source view Iw(x)I_w(x) by sampling pixel intensities from the target right view I2I_2 shifted along horizontal epipolar lines:

    Iw(x)=I2(x+D(x))=I2(x+fBd(x))I_w(x) = I_2\left(x + D(x)\right) = I_2\left(x + \frac{fB}{d(x)}\right)

    The network is trained end-to-end to minimize the photometric reconstruction error between IwI_w and I1I_1. At test time, the network operates purely as a single-image depth estimator, requiring only I1I_1 without stereo pairs or ground-truth depth supervision.

  2. Knowl 2 — Photometric Reconstruction and Disparity Smoothness Objective Function

    equation

    The end-to-end loss function for training the single-view depth prediction network over a dataset of NN stereo pairs {I1i,I2i}i=1N\{I_1^i, I_2^i\}_{i=1}^N is defined as:

    E=∑i=1N(Ereconsi+γEsmoothi)E = \sum_{i=1}^N \left( E_{\text{recons}}^i + \gamma E_{\text{smooth}}^i \right)

    where γ>0\gamma > 0 is a scalar hyperparameter controlling the regularization strength (set to γ=0.01\gamma = 0.01).

    The photometric reconstruction error EreconsiE_{\text{recons}}^i penalizes the intensity difference across the image spatial domain Ω⊂R2\Omega \subset \mathbb{R}^2 between the left image I1iI_1^i and the backward-warped right image IwiI_w^i:

    Ereconsi=∫Ω∥Iwi(x)−I1i(x)∥2 dx=∫Ω∥I2i(x+Di(x))−I1i(x)∥2 dxE_{\text{recons}}^i = \int_{\Omega} \|I_w^i(x) - I_1^i(x)\|^2 \, dx = \int_{\Omega} \left\| I_2^i\left(x + D^i(x)\right) - I_1^i(x) \right\|^2 \, dx

    where Di(x)=fB/di(x)D^i(x) = fB / d^i(x) is the horizontal disparity at pixel coordinate xx, ff is focal length, BB is the stereo baseline, and di(x)d^i(x) is the predicted depth.

    The smoothness prior EsmoothiE_{\text{smooth}}^i resolves the aperture problem and ambiguities in homogeneous photometric regions using an L2L_2 regularizer on the disparity gradient:

    Esmoothi=∫Ω∥∇Di(x)∥2 dxE_{\text{smooth}}^i = \int_{\Omega} \|\nabla D^i(x)\|^2 \, dx

  3. Knowl 3 — First-Order Taylor Linearization for Differentiable Backward Image Warping

    equation

    To compute gradients for standard backpropagation on the photometric reconstruction loss EreconsE_{\text{recons}}, the non-linear image warping function I2(x+Dn(x))I_2(x + D^n(x)) is linearized around the disparity estimate Dn−1(x)D^{n-1}(x) from iteration n−1n-1 using a first-order Taylor series expansion:

    I2(x+Dn(x))≈I2(x+Dn−1(x))+(Dn(x)−Dn−1(x))I2h(x+Dn−1(x))I_2\left(x + D^n(x)\right) \approx I_2\left(x + D^{n-1}(x)\right) + \left(D^n(x) - D^{n-1}(x)\right) I_{2h}\left(x + D^{n-1}(x)\right)

    where I2h(x)I_{2h}(x) denotes the horizontal spatial image gradient of I2I_2.

    This linearization is strictly valid under the assumption that the disparity update magnitude ∣Dn(x)−Dn−1(x)∣|D^n(x) - D^{n-1}(x)| remains small, which establishes the necessity for iterative coarse-to-fine multi-scale optimization to accurately recover large disparities (close distances).

  4. Knowl 4 — Fully Convolutional Depth Encoder Architecture with Skip Connections

    model/method

    The depth estimation network adapts the convolutional trunk of AlexNet up to layer C5, replacing the subsequent fully connected layers with a fully convolutional layer containing 2048 kernels of spatial size 5×55 \times 5, initialized to zero weights to yield initial zero-disparity predictions. This allows variable input sizes and spatial preservation.

    To recover spatial detail lost in deep pooling stages, a skip-connection architecture is used. Coarser disparity maps are upsampled via 4×44 \times 4 convolutional upsampling filters (initialized to bilinear interpolation). Features from earlier pooling layers (P3 and P2) are passed through 1×11 \times 1 convolutional layers (initialized with zero filters and biases) and added via element-wise summation to the upsampled disparity representations, producing progressively higher resolution disparity predictions across layers L8 through L12 (yielding resolutions 10×3710 \times 37, 22×7622 \times 76, 44×15244 \times 152, 88×30488 \times 304, and 176×608176 \times 608 from a 188×620188 \times 620 input image).

  5. Knowl 5 — Stage-Wise Coarse-to-Fine Training Algorithm

    algorithm

    To satisfy the small-motion requirement of the Taylor linearization across large disparities, the network is trained progressively across multiple resolutions:

    Input: Training stereo pairs {I_1^i, I_2^i}, number of scales S = 5 (L7 to L12)
    Output: Trained single-view depth estimation CNN
    Initialize AlexNet layers L1-L5 randomly; initialize 5x5 conv layer L7 with zeros
    Initialize base learning rate eta = 0.01, decay parameter alpha = 0.0005, gamma = 0.01
    for stage s = 1 to S do
        if s > 1 then
            Add upsampling deconvolution layer (initialized with bilinear kernel)
            if s in {2, 3} then
                Add skip connection from corresponding pooling layer (P3, P2) with 1x1 conv
            eta = eta / 4
        for epoch n = 1 to 100 do
            eta_n = eta / ((1 + alpha * n)^(n - 1))
            for each stereo pair (I_1, I_2) do
                Resize target I_2 and reference I_1 to current stage resolution
                Compute disparity map D = CNN(I_1)
                Compute warped target I_w using current disparity and linearization
                Compute loss E = E_recons + gamma * E_smooth
                Update CNN weights via SGD with momentum 0.9, weight decay 0.0005
    return CNN

  6. Knowl 6 — KITTI Benchmark Evaluation Setup and Error Metrics

    experimental setup

    Experiments are conducted on the outdoor KITTI dataset using the standard Eigen split (28 scenes for training, 28 for testing across "city", "residential", and "road" categories), comprising 22,600 stereo training pairs, 888 validation pairs, and 697 test images. Input left images are downsampled by a factor of 2 to 188×620188 \times 620, normalized to [−0.5,0.5][-0.5, 0.5]. Predicted disparities DD are upscaled to full resolution and mapped to depth via d=fB/Dd = fB/D with baseline B=0.54 mB = 0.54\,\text{m}. Predictions are clamped to [1,50] m[1, 50]\,\text{m} within the standard crop.

    Evaluation uses standard single-view depth metrics over valid ground-truth pixels TT:

    • Root Mean Squared Error (RMS): 1∣T∣∑i∈T∥di−digt∥2\sqrt{\frac{1}{|T|} \sum_{i \in T} \|d_i - d_i^{\text{gt}}\|^2}
    • Log RMS: 1∣T∣∑i∈T∥log⁡(di)−log⁡(digt)∥2\sqrt{\frac{1}{|T|} \sum_{i \in T} \|\log(d_i) - \log(d_i^{\text{gt}})\|^2}
    • Absolute Relative Error: 1∣T∣∑i∈T∣di−digt∣digt\frac{1}{|T|} \sum_{i \in T} \frac{|d_i - d_i^{\text{gt}}|}{d_i^{\text{gt}}}
    • Square Relative Error: 1∣T∣∑i∈T∥di−digt∥2digt\frac{1}{|T|} \sum_{i \in T} \frac{\|d_i - d_i^{\text{gt}}\|^2}{d_i^{\text{gt}}}
    • Threshold Accuracies: percentage of did_i satisfying max⁡(didigt,digtdi)=δ<thr\max\left(\frac{d_i}{d_i^{\text{gt}}}, \frac{d_i^{\text{gt}}}{d_i}\right) = \delta < \text{thr}, for thr∈{1.25,1.252,1.253}\text{thr} \in \{1.25, 1.25^2, 1.25^3\}.

  7. Knowl 7 — Performance Progression Across Upsampling Stages and Data Augmentation

    data/table

    The progression of single-view depth accuracy on the KITTI test set as resolution is increased stage-wise and with data augmentation is evaluated below:

    Methods Resolution RMS log RMS Abs. rel. Sq. rel. δ<1.25\delta < 1.25 δ<1.252\delta < 1.25^2 δ<1.253\delta < 1.25^3
    Ours L9 22×7622 \times 76 5.740 0.310 0.205 1.353 0.660 0.872 0.948
    Ours L10 + skip 46×15446 \times 154 5.850 0.338 0.246 1.673 0.607 0.842 0.937
    Ours L10 44×15244 \times 152 5.434 0.292 0.189 1.214 0.705 0.889 0.955
    Ours L11 88×30488 \times 304 5.326 0.285 0.179 1.177 0.721 0.892 0.958
    Ours L12 176×608176 \times 608 5.285 0.282 0.177 1.169 0.727 0.896 0.958
    Ours L12, Aug 8x 176×608176 \times 608 5.104 0.273 0.169 1.080 0.740 0.904 0.962

    These results show that progressively training higher-resolution upsampling layers consistently reduces RMS error from 5.740 down to 5.285. Adding a third skip connection at layer L10 degrades performance relative to simple upsampling without skip connection (RMS 5.850 vs. 5.434), indicating that early convolutional features capture photometric rather than depth properties. Applying 8×8\times data augmentation (random color scaling c∈[0.9,1.1]c \in [0.9, 1.1], spatial scale s∈[1,1.6]s \in [1, 1.6], and horizontal flip-and-swap) further reduces RMS to 5.104 and improves δ<1.25\delta < 1.25 accuracy to 0.740.

  8. Knowl 8 — Benchmarking Unsupervised Single-View Depth Estimation Against Supervised Methods

    data/table

    Comparison on the KITTI test set between the unsupervised autoencoder framework and supervised state-of-the-art single-view depth estimation approaches:

    Methods Resolution RMS log RMS Abs. rel. Sq. rel. δ<1.25\delta < 1.25 δ<1.252\delta < 1.25^2 δ<1.253\delta < 1.25^3
    Ours L12 176×608176 \times 608 5.285 0.282 0.177 1.169 0.727 0.896 0.958
    Ours L12, Aug 8x 176×608176 \times 608 5.104 0.273 0.169 1.080 0.740 0.904 0.962
    Mean – 9.635 0.444 0.412 5.712 0.556 0.752 0.870
    Make3D Dense 8.734 0.361 0.280 3.012 0.601 0.820 0.926
    Eigen et al. (coarse) 28×14428 \times 144 7.216 0.273 0.194 1.531 0.679 0.897 0.967
    Eigen et al. (fine) 27×14227 \times 142 7.156 0.270 0.190 1.515 0.692 0.899 0.967
    Fayao et al. (pt) superpix 7.421 – – – 0.613 0.858 0.949
    Fayao et al. (ft) superpix 7.046 – – – 0.656 0.881 0.958

    The unsupervised model, trained with randomly initialized weights and without depth supervision, achieves lower RMS (5.104 vs. 7.156) and lower absolute relative error (0.169 vs. 0.190) than the fully supervised multi-scale deep network of Eigen et al. initialized on ImageNet.

  9. Knowl 9 — Direct Photometric Autoencoding vs. Proxy Stereo Ground-Truth Supervision

    data/table

    Comparison on the KITTI test set between the proposed unsupervised autoencoder and supervised networks trained using proxy depth labels generated by off-the-shelf stereo algorithms:

    Methods Coverage RMS log RMS Abs. rel. Sq. rel. δ<1.25\delta < 1.25 δ<1.252\delta < 1.25^2 δ<1.253\delta < 1.25^3
    Ours L12 100% 5.285 0.282 0.177 1.169 0.727 0.896 0.958
    HS →\to CNN, γ=0.01\gamma = 0.01 100% 6.691 0.385 0.309 2.657 0.476 0.750 0.891
    HS →\to CNN 100% 6.292 0.338 0.238 1.639 0.573 0.841 0.941
    SGM →\to CNN 100% 5.680 0.300 0.185 1.370 0.703 0.886 0.955
    HS-Stereo, γ=0.01\gamma = 0.01 100% 6.077 0.381 0.299 3.264 0.677 0.822 0.900
    HS-Stereo 100% 6.760 0.366 0.254 4.040 0.754 0.872 0.928
    SGM-Stereo 87% 3.030 0.150 0.064 0.506 0.955 0.979 0.989

    Training a CNN with least-squares loss to predict disparities estimated by Horn-Schunck (HS →\to CNN, RMS 6.292) or Semi-Global Matching (SGM →\to CNN, RMS 5.680) performs worse than direct end-to-end photometric autoencoding (RMS 5.285). Supervised training on proxy stereo maps leads the network to internalize the systematic failure modes and stereo artifacts (such as reconstruction holes in SGM) of the generating algorithm.

  10. Knowl 10 — Limitations of Disparity Linearization and Bilinear Upsampling

    limitation

    The framework exhibits two primary limitations:

    1. The reliance on a first-order Taylor expansion to linearize image warping requires that disparity increments between iterations remain small. This restricts single-stage end-to-end optimization and necessitates a complex multi-stage coarse-to-fine schedule to resolve large disparities (objects close to the camera).
    2. Using simple bilinear interpolation and L2L_2 smoothness regularization leads to oversmoothed predictions and blurred depth boundaries at high resolutions, failing to capture sharp object silhouettes without post-processing mechanisms like fully connected CRFs or atrous convolutions.

Coverage note — None was omitted; all key contributions including the geometric autoencoder formulation, loss formulation, linearization, multi-scale network architecture, training algorithm, and experimental benchmarks were captured.

References

  1. 1.Agrawal, P., Carreira, J., Malik, J.: Learning to see by moving. In: IEEE International Conference on Computer Vision (ICCV) (2015)
  2. 2.Brox, T., Bruhn, A., Papenberg, N., Weickert, J.: High accuracy optical flow estimation based on a theory for warping. In: Pajdla, T., Matas, J.G. (eds.) ECCV 2004. LNCS, vol. 3024, pp. 25–36. Springer, Heidelberg (2004)
  3. 3.Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected CRFs. In: International Conference on Learning Representations (ICLR) (2015)
  4. 4.Chen, Z., Sun, X., Wang, L., Yu, Y., Huang, C.: A deep visual correspondence embedding model for stereo matching costs. In: IEEE International Conference on Computer Vision (ICCV) (2015)
  5. 5.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: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)
  6. 6.Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: a large-scale hierarchical image database. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2009)
  7. 7.Dosovitskiy, A., Springenberg, J.T., Riedmiller, M., Brox, T.: Discriminative unsupervised feature learning with convolutional neural networks. In: Advances in Neural Information Processing Systems (NIPS) (2014)
  8. 8.Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. In: Advances in Neural Information Processing Systems (NIPS) (2014)
  9. 9.Fan, X., Zheng, K., Lin, Y., Wang, S.: Combining local appearance and holistic view: dual-source deep neural networks for human pose estimation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015)
  10. 10.Flynn, J., Neulander, I., Philbin, J., Snavely, N.: Deepstereo: learning to predict new views from the world’s imagery (2016)
  11. 11.Garg, R., Pizarro, L., Rueckert, D., Agapito, L.: Dense multi-frame optic flow for non-rigid objects using subspace constraints. In: Kimmel, R., Klette, R., Sugimoto, A. (eds.) ACCV 2010, Part IV. LNCS, vol. 6495, pp. 460–473. Springer, Heidelberg (2011)
  12. 12.Garg, R., Roussos, A., Agapito, L.: Dense variational reconstruction of non-rigid surfaces from monocular video. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2013)
  13. 13.Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: the KITTI dataset. Int. J. Robot. Res. (IJRR) 32, 1229–1235 (2013)
  14. 14.Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)
  15. 15.Handa, A., Patraucean, V., Badrinarayanan, V., Stent, S., Cipolla, R.: Scenenet: understanding real world indoor scenes with synthetic data. arXiv preprint (2015). arXiv:1511.07041
  16. 16.Hirschmuller, H.: Accurate and efficient stereo processing by semi-global matching and mutual information. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2005)
  17. 17.Horn, B.K., Schunck, B.G.: Determining optical flow. In: 1981 technical symposium east, pp. 319–331. International Society for Optics and Photonics (1981)
  18. 18.Koltun, V.: Efficient inference in fully connected CRFs with Gaussian edge potentials. In: Neural Information Processing Systems (NIPS) (2011)
  19. 19.Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems (NIPS) (2012)
  20. 20.Vijay Kumar, B.G., Carneiro, G., Reid, I.: Learning local image descriptors with deep siamese and triplet convolutional networks by minimising global loss functions. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)
  21. 21.Ladicky, L., Shi, J., Pollefeys, M.: Pulling things out of perspective. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)
  22. 22.Li, B., Shen, C., Dai, Y., van den Hengel, A., He, M.: Depth and surface normal estimation from monocular images using regression on deep features and hierarchical CRFs. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015)
  23. 23.Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014, Part V. LNCS, vol. 8693, pp. 740–755. Springer, Heidelberg (2014)
  24. 24.Liu, F., Shen, C., Lin, G., Reid, I.: Learning depth from single monocular images using deep convolutional neural fields. In: IEEE Transactions on Pattern Analysis and Machine Intelligence (2016)
  25. 25.Long, G., Kneip, L., Alvarez, J.M., Li, H.: Learning image matching by simply watching video. CoRR abs/1603.06041 (2016). http://arxiv.org/abs/1603.06041
  26. 26.Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015)
  27. 27.Newcombe, R.A., Lovegrove, S.J., Davison, A.J.: Dtam: dense tracking and mapping in real-time. In: IEEE International Conference on Computer Vision (ICCV) (2011)
  28. 28.Noh, H., Hong, S., Han, B.: Learning deconvolution network for semantic segmentation. In: IEEE International on Computer Vision (ICCV) (2015)
  29. 29.Saxena, A., Sun, M., Ng, A.: Make3d: learning 3d scene structure from a single still image. IEEE Trans. Pattern Anal. Mach. Intell. (PAMI) 31, 824–840 (2009)
  30. 30.Steinbrücker, F., Pock, T., Cremers, D.: Large displacement optical flow computation withoutwarping. In: IEEE International Conference on Computer Vision (ICCV) (2009)
  31. 31.Vedaldi, A., Lenc, K.: Matconvnet - convolutional neural networks for matlab (2015)
  32. 32.Xie, J., Girshick, R., Farhadi, A.: Deep. 3d: fully automatic 2d-to-3d video conversion with deep convolutional neural networks. arXiv preprint (2016). arXiv:1604.03650
  33. 33.Zach, C., Pock, T., Bischof, H.: A duality based approach for realtime TV-L1^1 optical flow. In: Hamprecht, F.A., Schnörr, C., Jähne, B. (eds.) Pattern Recognition, vol. 4713, pp. 214–223. Springer, Heidelberg (2007)
  34. 34.Zbontar, J., LeCun, Y.: Computing the stereo matching cost with a convolutional neural network. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015)
  35. 35.Zheng, S., Jayasumana, S., Romera-Paredes, B., Vineet, V., Su, Z., Du, D., Huang, C., Torr, P.H.S.: Conditional random fields as recurrent neural networks. In: IEEE International Conference on Computer Vision (ICCV) (2015)

Citation

MLA
Garg, R., et al. “Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue”. arXiv, 2016, http://arxiv.org/abs/1603.04992v2.
APA
Garg, R., BG, V. K., Carneiro, G., & Reid, I. (2016). Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue. arXiv. http://arxiv.org/abs/1603.04992v2
Chicago
Garg, R., V. K. BG, G. Carneiro, and I. Reid. 2016. “Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue”. arXiv. http://arxiv.org/abs/1603.04992v2.
Harvard
Garg, R. et al. (2016) “Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1603.04992v2.
Vancouver
1. Garg R, BG VK, Carneiro G, Reid I (2016) Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue. arXiv

BibTeX

@article{garg2016unsupervised,
  title = {Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue},
  author = {Garg, Ravi and BG, Vijay Kumar and Carneiro, Gustavo and Reid, Ian},
  year = {2016},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1603.04992v2},
  eprint = {1603.04992}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF