Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

DiverseDepth Project

This project aims to improve the generalization ability of the monocular depth estimation method on diverse scenes. We propose a learning method and a diverse dataset, termed DiverseDepth, to solve this problem. The DiverseDepth contents have been published in our "Virtual Normal" TPAMI version.

This repository contains the source code of our paper (the DiverseDepth part):

  1. Wei Yin, Yfan Liu, Chunhua Shen, Virtual Normal: Enforcing Geometric Constraints for Accurate and Robust Depth Prediction.
  2. Wei Yin, Xinlong Wang, Chunhua Shen, Yifan Liu, Zhi Tian, Songcen Xu, Changming Sun, Dou Renyin. DiverseDepth: Affine-invariant Depth Prediction Using Diverse Data

Training codes have been released!!

Some Results

Any images online Point cloud

Some Dataset Examples



  • Generalization: Our method demonstrates strong generalization ability on several zero-shot datasets. The predicted depth is affine-invairiant.



We collect multi-source data to construct our DiverseDepth dataset. The It consists of three parts: Part-in (collected from taskonomy): contains over 100K images Part-out (collected from DIML, we have reprocessed its disparity): contains over 120K images Part-fore (collected from webstereo images and videos): contains 109703 images. We used the GNet method to recompute the disparity of DIML data instead of original provided disparity maps. We provide two ways to download data.

  1. Download from Cloudstor. You can download them with the following method.
  1. Download from Google Drive. See here. for details.

Quick Start (Inference)

  1. Download the model weights

  2. Prepare data.

    • Move the downloaded weights to <project_dir>/
    • Put the testing RGB images to <project_dir>/Minist_Test/test_images/. Predicted depths and reconstructed point cloud are saved under <project_dir>/Minist_Test/test_images/outputs
  3. Test monocular depth prediction. Note that the predicted depths are affine-invariant.

export PYTHONPATH="<PATH to DiverseDepth>"
# run the ResNet-50
python ./Minist_Test/tools/ --load_ckpt model.pth


  1. Download the ResNeXt pretrained weight and put it under Train/datasets/resnext_pretrain
  2. Download the training data. Refer to ''. All data are organized under the Train/datasets. The structure of all data are as follows.
|    |--DiverseDepth
|    |   |--annotations
|    |   |--depths
|    |   |--rgbs
|    |--taskonomy
|    |   |--annotations
|    |   |--depths
|    |   |--rgbs
|    |   |--ins_planes
|    |--DIML_GANet
|    |   |--annotations
|    |   |--depth
|    |   |--rgb
|    |   |--sky_mask
|    |--resnext_pretrain
|    |   |--resnext50_32x4d.pth
  1. Train the network. The default setting used 4 gpus. If you want to use more gpus, please set $CUDA_VISIBLE_DEVICES, such as export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7. The --batchsize is the number of samples on a single gpu.

    cd Train/scripts
  2. Test the network on a benchmark. We provide a sample code for testing on NYU. Please download the NYU testing data test.mat for evaluation. If you want to test on other benchmarks, you can follow the sample code.

    cd Train/scripts


  title={Virtual Normal: Enforcing Geometric Constraints for Accurate and Robust Depth Prediction},
  author={Yin, Wei and Liu, Yifan and Shen, Chunhua},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},


Wei Yin: