MonoDepth-FPN-PyTorch

A simple end-to-end model that achieves state-of-the-art performance in depth prediction implemented in PyTorch. We used a Feature Pyramid Network (FPN) backbone to estimate depth map from a single input RGB image. We tested the performance of our model on the NYU Depth V2 Dataset (Official Split) and the KITTI Dataset (Eigen Split).

Requirements

• Python 3
• Jupyter Notebook (for visualization)
• PyTorch
  • Tested with PyTorch 0.3.0.post4
• CUDA 8 (if using CUDA)

To Run

python3 main_fpn.py --cuda --bs 6

To continue training from a saved model, use

python3 main_fpn.py --cuda --bs 6 --r True --checkepoch 10

To visualize the reconstructed data, run the jupyter notebook in vis.ipynb.

Data Processing

NYU Depth V2 Dataset

• The NYU Depth V2 dataset contains a variety of indoor scenes, with 249 scenes for training and 215 scenes for testing. We used the official split for training and testing. This github page provides a very detailed walkthrough and matlab code for data processing.
• Following previous works, we used the official toolbox which uses the Colorization method proposed by Levin et al. to fill in the missing values of depth map in the training set.
• To make comparison with previous works, we evaluated our model using the official evaluation set of 654 densely labeled image pairs.
• The images and depth maps in the NYU Depth V2 dataset are both of size 640x480. During training, we loaded the images as-is, and downscaled the depth maps to 160x120. The proposed model produces depth maps of size 160x120, and upsampled to the original size for evaluation.
• We employed random crop, random rotate of 10 degrees, color jitter of brightness, contrast, saturation and hue of variation 0.1. To save time during training, we performed data augmentation in advance by running dataset/augment.py.
• PyTorch does not support transformation for both the input and the target, so we implemented joint transforms for data augmentation.

KITTI Dataset

• The KITTI dataset consists of 61 outdoor scenes with “city”, “road”, and “residential” categories. Following the official tutorial, we got about 22k image and depth map pairs in the training dataset.
• Following previous works, we used the same NYU tool box to fill in the missing values of the sparse depth maps in the training set.
• To compare with the performances of previous studies, we evaluate on the test split of KITTI dataset proposed by Eigen et al.
• The images and depth maps in the KITTI dataset are both of size about 1280x384. During training, we downscaled the images to size 640x192, and downscaled the depth maps to 160x48. The proposed model produces depth maps of size 160x48, and upsampled to the original size for evaluation.
• For better visualization in quantitative evaluation, we filled in the missing depth value in the ground truth depth map.

Model

• We used Feature Pyramid Network (FPN) with ResNet101 as backbone (shaded yellow), loaded ImageNet pretrained weight as weight initialization.
• We used pixel-shuffle for upsampling and fuse feature maps with add operation; bilinear interpolation is employed after pixel-shuffle to deal with inconsistent feature map size.
• Two consecutive 3x3 convolutions for feature processing.
• No non-linearity in the top-down branch of FPN, and ReLU in other convolution layers, and Sigmoid in the prediction layer for better stability.
• Trained on the weighted sum of the depth loss, the gradient loss, and the normal loss for 20 epochs for the NYU Depth V2 dataset, and 40 epochs for the KITTI dataset; gradient loss added after epoch 1, and normal loss added after epoch 10.
• Outputs prediction of size ¼, and evaluated after bilinear upsampling.

Loss Function

We employed three parts in the loss function in our model. The loss is a weighted sum of 3 parts: the depth loss, the gradient loss and the surface normal loss.

Depth Loss

The depth loss is RMSE in log scale, which we found converges better than L1 and L2 norm. Supervising in log scale makes the classifier focus more on closer objects.

Gradient Loss

The gradient of depth maps is obtained by a Sobel filter; the gradient loss is the L1 norm of the difference.

Surface Normal Loss

We also employed the normal vector loss proposed by Hu et al., which helps refining details.

The weight ratio between the three loss was set to 1:1:1.

Qualitative Evaluation

KITTI

Comparison with state-of-the-art methods:

More comparison:

Quantitative Evaluation

KITTI

We use the following depth evaluation metrics used by Eigen et al.:

where T is the number of valid pixels in the test set.

Discussion

• FPN is an effective backbone for monocular depth estimation because of its ability to extract features and semantics at different scales. It can achieve its potential if guided by proper loss functions.
• Gradient and normal losses help prevent the model getting stuck in local optimum and guide it toward better convergence, as shown in the ablation study.
• Some existing state-of-the-art methods ourperform ours in certain metrics, and we believe that this is due to the adaptive BerHu loss.

Related Work

• Eigen et al. were the first to use CNNs for depth estimation: predicting a coarse global output and then a finer local network.
• Laina et al. explored CNNs pre-trained with AlexNet, VGG16 and ResNet.
• Godard et al. and Kuznietsov et al. introduced unsupervised and semi-supervised methods that relies on 3D geometric assumptions of left-right consistency. They trained the network using binocular image data.
• Fu et al. proposed a framework based on classification of discretized depth ranges and regression. They supervised depth prediction at different resolutions and used a fusion network to produce the final output.
• Hu et al. proposed a novel loss of normal vectors in addition to the conventional depth and gradient losses. They first trained a base network without skip connections, and then train the refine network using the novel loss after freezing the weights of the base network.
• This project: Fully convolutional with no FC layers needed, FPN provides a straightforward and unified backbone to extract features maps that incorporates strong and localized semantic information. We employ an easy to follow curriculum to add in gradient loss and normal loss during training, all losses are calculated on the output feature maps instead of intermediate ones.

References

• Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multiscale deep network. In: Advances in neural information processing systems (NIPS). (2014) 2366–2374
• Fu, Huan, Mingming Gong, Chaohui Wang and Dacheng Tao. “A Compromise Principle in Deep Monocular Depth Estimation.” CoRR abs/1708.08267 (2017): n. pag.
• R. Garg, V. Kumar, G. Carneiro, and I. Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Proc. of the European Conf. on Computer Vision (ECCV), 2016.
• Geiger, Andreas, et al. "Vision meets robotics: The KITTI dataset." The International Journal of Robotics Research 32.11 (2013): 1231-1237.
• C. Godard, O. Mac Aodha, and G. J. Brostow. Unsupervised monocular depth estimation with left-right consistency. arXiv:1609.03677v2, 2016.
• Hu, Junjie, et al. "Revisiting Single Image Depth Estimation: Toward Higher Resolution Maps with Accurate Object Boundaries." arXiv preprint arXiv:1803.08673 (2018).
• Kuznietsov, Yevhen, Jörg Stückler, and Bastian Leibe. "Semi-supervised deep learning for monocular depth map prediction." Proc. of the IEEE Conference on Computer Vision and Pattern Recognition. 2017.
• I. Laina, C. Rupprecht, V. Belagiannis, F. Tombari, and N. Navab. Deeper depth prediction with fully convolutional residual networks. In Proc. of the Int. Conf. on 3D Vision (3DV), 2016.
• Levin, Anat, Dani Lischinski, and Yair Weiss. "Colorization using optimization." ACM Transactions on Graphics (ToG). Vol. 23. No. 3. ACM, 2004.
• Silberman, Nathan, et al. "Indoor segmentation and support inference from rgbd images." European Conference on Computer Vision. Springer, Berlin, Heidelberg, 2012.