Home

Awesome

On the uncertainty of <br> self-supervised monocular depth estimation

Demo code of "On the uncertainty of self-supervised monocular depth estimation", Matteo Poggi, Filippo Aleotti, Fabio Tosi and Stefano Mattoccia, CVPR 2020.

At the moment, we do not plan to release training code.

[Paper] - [Poster] - [Youtube Video]

<p align="center"> <img src=https://mattpoggi.github.io/assets/img/uncertainty/poggi2020cvpr.gif> </p>

Citation

@inproceedings{Poggi_CVPR_2020,
  title     = {On the uncertainty of self-supervised monocular depth estimation},
  author    = {Poggi, Matteo and
               Aleotti, Filippo and
               Tosi, Fabio and
               Mattoccia, Stefano},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2020}
}

Contents

  1. Abstract
  2. Usage
  3. Contacts
  4. Acknowledgements

Abstract

Self-supervised paradigms for monocular depth estimation are very appealing since they do not require ground truth annotations at all. Despite the astonishing results yielded by such methodologies, learning to reason about the uncertainty of the estimated depth maps is of paramount importance for practical applications, yet uncharted in the literature. Purposely, we explore for the first time how to estimate the uncertainty for this task and how this affects depth accuracy, proposing a novel peculiar technique specifically designed for self-supervised approaches. On the standard KITTI dataset, we exhaustively assess the performance of each method with different self-supervised paradigms. Such evaluation highlights that our proposal i) always improves depth accuracy significantly and ii) yields state-of-the-art results concerning uncertainty estimation when training on sequences and competitive results uniquely deploying stereo pairs.

Usage

Requirements

Getting started

Clone Monodepth2 repository and set it up using

sh prepare_monodepth2_engine.sh

Download KITTI raw dataset and accurate ground truth maps

sh prepare_kitti_data.sh kitti_data

with kitti_data being the datapath for the raw KITTI dataset. The script checks if you already have raw KITTI images and ground truth maps there. Then, it exports ground truth depths according to Monodepth2 format.

Pretrained models

You can download the following pre-trained models:

Run inference

Launch variants of the following command (see batch_generate.sh for a complete list)

python generate_maps.py --data_path kitti_data \
                        --load_weights_folder weights/M/Monodepth2-Post/models/weights_19/ \
                        --post_process \
                        --eval_split eigen_benchmark \
                        --output_dir experiments/Post/ \
                        --eval_mono

It assumes you have downloaded pre-trained models and placed them in the weights folder. Use --eval_stereo for S and MS models.

Extended options (in addition to Monodepth2 arguments):

Results are saved in --output_dir/raw and are ready for evaluation. Qualitatives are saved in --output_dir/qual.

Run evaluation

Launch the following command

python evaluate.py --ext_disp_to_eval experiments/Post/raw/ \
                   --eval_mono \
                   --max_depth 80 \
                   --eval_split eigen_benchmark \
                   --eval_uncert

Optional arguments:

Results

Results for evaluating Post depth and uncertainty maps:

   abs_rel |   sq_rel |     rmse | rmse_log |       a1 |       a2 |       a3 |
&   0.088  &   0.508  &   3.842  &   0.134  &   0.917  &   0.983  &   0.995  \\

   abs_rel |          |     rmse |          |       a1 |          |
      AUSE |     AURG |     AUSE |     AURG |     AUSE |     AURG |
&   0.044  &   0.012  &   2.864  &   0.412  &   0.056  &   0.022  \\

Minor changes can occur with different versions of the python packages (not greater than 0.01)

Minor differences from the paper

Contacts

m [dot] poggi [at] unibo [dot] it

Acknowledgements

Thanks to Niantic and Clément Godard for sharing Monodepth2 code