Home

Awesome

Monodepth uncertainty

project webpage

<img src="./figs/multi-modal.png" width="600">

Requirements

Weights

Test

On KITTI (test set)

CUDA_VISIBLE_DEVICES=3 python test.py --check weights-kitti-nyu-resizespp-100-v2/model-320000 --con configs/model-1s100.config --input_image figs/kitti_2011_09_26_drive_0001_sync_02_0000000012.jpg --max_depth 100

On NYU (test set)

CUDA_VISIBLE_DEVICES=3 python test.py --check weights-kitti-nyu-resizespp-100-v2/model-320000 --con configs/model-1s100.config --input_image figs/nyu_1449.jpg --max_depth 10

KITTI-expectation

KITTI expected depth

KITTI-entropy

KITTI entropy

NYU-expectation

NYU expected depth

NYU-entropy

NYU entropy

Training

If you are interested in training the binary depth estimator, see the tf code for training. See README for details to train the model. Note the training code is very messy. It is recommended to start from Monodepth and use our code as a reference to modify the dataloader as well as the loss functions.

Citation

@InProceedings{yang2019inferring,
author = {Yang, Gengshan and Hu, Peiyun and Ramanan, Deva},
title = {Inferring distributions over depth from a single image},
booktitle = {2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year = {2019}
}