DORN_depth_estimation_Pytorch

This is an unoficial Pytorch implementation of Deep Ordinal Regression Network for Monocular Depth Estimation paper by Fu et. al.

Table. Performance on NYU V2.

Source	δ1	δ2	δ3	rel	log10	rms
Original paper*	0.828	0.965	0.992	0.115	0.051	0.509
This repo*	0.806	0.957	0.989	0.151	0.062	0.586

*Note, that the data splits are different (see Known Differences below for details). The worse performance might be due to the smaller training set (795 vs about 120K images).

How to use

These steps show how to run the code on the official split of the NYU V2 depth dataset.

To prepare data:

• Download nyu_depth_v2_labeled.mat (data) and splits.mat.
• Edit create_nyu_h5.py to add data_path (folder with the .mat files from previous step) and output_path.
• Run:

python create_nyu_h5.py

For start training on NYU V2 run:

train.py [-h] [--dataset DATASET] [--data-path DATA_PATH]
              [--pretrained] [--epochs EPOCHS] [--bs BS] [--bs-test BS_TEST]
              [--lr LR] [--gpu GPU]

Or simply:

python train.py --data-path DATA_PATH --pretrained

(where DATA_PATH is same as output_path used during preparing data).

For more info on arguments run:

python train.py --help

To train on a different dataset, implementation of the DataLoader is required.

To monitor training, use Tensorboard:

tensorboard --logdir ./logs/

Known Differences

The implementation closely follows the paper and the official repo with some exceptions. The list of known differences:

• Only training on the labeled part of NYU V2 is currently implemented (not on all the raw data).
• ColorJitter is used instead of the color transformation from the Eigen's paper.
• Feature extractor is pretrained on a different dataset.

Pretrained feature extractor

DORN uses a modified version of ResNet-101 as a feature extractor (with dilations and three 3x3 convolutional layers in the begining instead of one 7x7 layer). If you select pretrained=True, weights pretrained on MIT ADE20K dataset will be loaded from this project. This is different from the paper (the authors suggest pretraining on ImageNet). That is the only suitable pretrained model on the Web that I am aware of.

Requirements

• Python 3
• Pytorch (version 1.3 tested)
• Torchvision
• Tensorboard

Acknowledgements

The code is based on this implementation.