Skip to content
Marr Revisited: 2D-3D Alignment via Surface Normal Prediction
Branch: master
Clone or download
Latest commit 13ec38f Dec 26, 2016
Type Name Latest commit message Commit time
Failed to load latest commit information.
caffe @ 3222608 updated instruction Sep 24, 2016
normals updated code Dec 25, 2016
.gitmodules first commit Jul 25, 2016 first commit Jul 25, 2016 updated instructions Oct 22, 2016

Marr Revisited: 2D-3D Alignment via Surface Normal Prediction

Created by Aayush Bansal, Bryan Russell, and Abhinav Gupta at CMU and Adobe Research.

Acknowledgements: a huge thanks to Xinlei Chen for all discussions and help.


This work is a part of our CVPR-2016 paper on 2D-3D model alignment via surface normal prediction. In this work, we predict detailed surface normals through the non-linear optimization of sparse hypercolumn features. The idea of sparsely sampling hypercolumn features allowed the non-linear optimization by a multi-layer perceptron. This work is the current state-of-the-art (as of July 2016) on NYUD dataset for both global and local layout. For more details, refer to the text in our paper.

The codes can be downloaded using the following command:

git clone --recursive
cd MarrRevisited

Here, we give step-by-step instruction to use our code for training models and estimating surface normal maps. We will release the codes for pose estimation soon, and optimized version of surface normals sometime soon.

Installing the Code

  1. Installing Caffe Toolbox - We have included a pointer to Caffe as a submodule. Our required layers are available within this submodule. You will need to pull it using the following command (in case you want to use it for some other project):

       git clone

    To install Caffe, please follow the instructions on their project page.

  2. The details corresponding to surface normal estimation is present in the "normals/" directory.

     cd normals

    The prototxt files for training the model and predicting the surface normal map from a single 2D image are available in "normals/net" directory:

      ls net

    There are two folders - a. conv; b. deconv. In our paper, we used the ones in 'a. conv'.

    The required caffe models (VGG-16 model converted to a fully convolutional model, and our best model for surface normal estimation) can be downloaded from :

    # download the models for estimating surface normal maps or to train your own model.
    cd cachedir
    tar -xvzf surface_normal_models.tar.gz
    cd ..

Specific instructions are given below to use our code for estimating surface normal maps from our trained model, and how to train a new model using NYU-v2 depth dataset.


You can run the demo code inside MATLAB (assuming you are in the "normals/" directory):


The results can be seen here :

   ls ./cachedir/demo_results/

Evaluation Script

The evaluation script is in "normals/eval" directory. Our results can be downloaded using following (given you are in "normals/" directory):

# The surface normal maps generated using our model.
cd cachedir/
mv results best_model  
cd ..

The other thing needed for evaluation are the surface normal maps from Kinect data on test set. Please see the section below to download the data. Once you have the required data, run following inside MATLAB:


Training your own model for surface normal prediction on NYUv2

In this work, we showed evaluation on NYU-v2 depth dataset. We now describe how to use our code with this dataset:

Training a model -

  1. Download the NYU-v2 dataset:
# 1449 labeled examples
mkdir dataset
cd dataset
cd ..
# Raw video frames
cd dataset
cd ..

The train, val, and test set splits can be downloaded using:

# train-val-test splits
cd dataset
tar -xvzf NYU_splits.tar.gz
cd ..

Note that Silberman et al. (ECCV 2012) contained the train-test split. Gupta et al. (ECCV 2014) provided an extra train-val split from training split. We use the same splits (also provided above).

  1. Surface Normal from Kinect - We used the surface normal provided by Ladicky et al. (ECCV 2014) for 1449 images. Those normal maps can be downloaded using:

    # Surface normal maps from kinect data for 1449 images.
    cd dataset
    cd ..
  2. We used the code from Wang et al. (CVPR 2015) for raw video frames as it was computationally too expensive to compute the normals using the approach of Ladicky et al. on all video frames. The codes can be downloaded using

    # Code to extract surface normal maps from kinect using Wang et al. (CVPR 2015)
    # Details to use this code are given in the folder
    cd toolbox/
    tar -xvf kinect_normals_code.tar.gz
    cd ..
    # For more details about this code, see the script getNormals.m
    # All the required data is present in this folder

    Note that in our work, we ignored the pixels belonging to invalid depth data. Therefore, other than the normal maps and images, one requires a map of valid pixels for each image.

  3. Data Preparation - Using the dataset, one can prepare the required data to train a new model for surface normal estimation. As mentioned in our paper, we resized the images to 224x224. Similarly we resized the normals to 224x224 and then normalized them. The valid pixel depth map is also resized to same value. To provide zero padding, we padded the images by 100 pixels.

  4. Once you have corresponding training images, normal maps, and valid depth data information (provided that you have successfully installed the code and downloaded the models) -- modify the paths mentioned in scripts and train.prototxt in the folder "normals/net/conv" to refer to data location. After these changes, simply use to train our model.

    # Training a new model after making required modifications
    cd net/conv
    # We ran the network for 60K iterations. This is roughly 1 epoch of training data.

Using the trained model

Once you have successfully trained your model, it can be easily used to compute normal maps using the deploy file. See the demo code in demo/.


Please contact Aayush Bansal in case of any query or comment about the code.


This code is copyrighted by the authors, CMU, and Adobe, and is for non-commercial research purposes only. Please contact the authors, CMU, or Adobe if you are interested in licensing for commercial purposes.


Please cite our paper if you use this code for any publishable academic research:

  title={Marr {R}evisited: 2{D}-3{D} Model Alignment via Surface Normal Prediction},
  author={Bansal, Aayush and Russell, Bryan and Gupta, Abhinav},


More details about our network architecture are available in PixelNet paper:

  title={PixelNet: {T}owards a {G}eneral {P}ixel-level {A}rchitecture},
  author={Bansal, Aayush and Chen, Xinlei, and  Russell, Bryan and Gupta, Abhinav and Ramanan, Deva},
  Journal={arXiv preprint arXiv:1609.06694},
You can’t perform that action at this time.