Home

Awesome

Video Autoencoder: self-supervised disentanglement of 3D structure and motion

This repository contains the code (in PyTorch) for the model introduced in the following paper:

Video Autoencoder: self-supervised disentanglement of 3D structure and motion
Zihang Lai, Sifei Liu, Alexi A. Efros, Xiaolong Wang
ICCV, 2021
[Paper] [Project Page] [12-min oral pres. video] [3-min supplemental video]

Figure

Citation

@inproceedings{Lai21a,
        title={Video Autoencoder: self-supervised disentanglement of 3D structure and motion},
        author={Lai, Zihang and Liu, Sifei and Efros, Alexei A and Wang, Xiaolong},
        booktitle={ICCV},
        year={2021}
}

Contents

  1. Introduction
  2. Data preparation
  3. Training
  4. Evaluation
  5. Pretrained model

Introduction

Figure We present Video Autoencoder for learning disentangled representations of 3D structure and camera pose from videos in a self-supervised manner. Relying on temporal continuity in videos, our work assumes that the 3D scene structure in nearby video frames remains static. Given a sequence of video frames as input, the Video Autoencoder extracts a disentangled representation of the scene including: (i) a temporally-consistent deep voxel feature to represent the 3D structure and (ii) a 3D trajectory of camera poses for each frame. These two representations will then be re-entangled for rendering the input video frames. Video Autoencoder can be trained directly using a pixel reconstruction loss, without any ground truth 3D or camera pose annotations. The disentangled representation can be applied to a range of tasks, including novel view synthesis, camera pose estimation, and video generation by motion following. We evaluate our method on several large-scale natural video datasets, and show generalization results on out-of-domain images.

Dependencies

The following dependencies are not strict - they are the versions that we use.

Data preparation

RealEstate10K:

  1. Download the dataset from RealEstate10K.
  2. Download videos from RealEstate10K dataset, decode videos into frames. You might find the RealEstate10K_Downloader written by cashiwamochi helpful. Organize the data files into the following structure:
RealEstate10K/
    train/
        0000cc6d8b108390.txt
        00028da87cc5a4c4.txt
        ...
    test/
        000c3ab189999a83.txt
        000db54a47bd43fe.txt
        ...
dataset/
    train/
        0000cc6d8b108390/
            52553000.jpg
            52586000.jpg
            ...
        00028da87cc5a4c4/
            ...
    test/
        000c3ab189999a83/
        ...
  1. Subsample the training set at one-third of the original frame-rate (so that the motion is sufficiently large). You can use scripts/subsample_dataset.py.
  2. A list of videos ids that we used (10K for training and 5K for testing) is provided here:
    1. Training video ids and testing video ids.
    2. Note: as time changes, the availability of videos could change.

Matterport 3D (this could be tricky):

  1. Install habitat-api and habitat-sim. You need to use the following repo version (see this SynSin issue for details):

    1. habitat-sim: d383c2011bf1baab2ce7b3cd40aea573ad2ddf71
    2. habitat-api: e94e6f3953fcfba4c29ee30f65baa52d6cea716e

  2. Download the models from the Matterport3D dataset and the point nav datasets. You should have a dataset folder with the following data structure:

    root_folder/
     mp3d/
         17DRP5sb8fy/
             17DRP5sb8fy.glb  
             17DRP5sb8fy.house  
             17DRP5sb8fy.navmesh  
             17DRP5sb8fy_semantic.ply
         1LXtFkjw3qL/
             ...
         1pXnuDYAj8r/
             ...
         ...
     pointnav/
         mp3d/
             ...

  3. Walk-through videos for pretraining: We use a ShortestPathFollower function provided by the Habitat navigation package to generate episodes of tours of the rooms. See scripts/generate_matterport3d_videos.py for details.

  4. Training and testing view synthesis pairs: we generally follow the same steps as the SynSin data instruction. The main difference is that we precompute all the image pairs. See scripts/generate_matterport3d_train_image_pairs.py and scripts/generate_matterport3d_test_image_pairs.py for details.

Replica:

  1. Testing view synthesis pairs: This procedure is similar to step 4 in Matterport3D - with only the specific dataset changed. See scripts/generate_replica_test_image_pairs.py for details.

Configurations

Finally, change the data paths in configs/dataset.yaml to your data location.

Pre-trained models

Training:

Use this script:

CUDA_VISIBLE_DEVICES=0,1 python train.py --savepath log/train --dataset RealEstate10K

Some optional commands (w/ default value in square bracket):

Evaluation:

1. Generate test results:

Use this script (for testing RealEstate10K):

CUDA_VISIBLE_DEVICES=0 python test_re10k.py --savepath log/model --resume log/model/checkpoint.tar --dataset RealEstate10K

or this script (for testing Matterport3D/Replica):

CUDA_VISIBLE_DEVICES=0 python test_mp3d.py --savepath log/model --resume log/model/checkpoint.tar --dataset Matterport3D

Some optional commands:

Running this will generate a output folder where the results (videos and poses) save. If you want to visualize the pose, use packages for evaluation of odometry, such as evo. If you want to quantitatively evaluate the results, see 2.1, 2.2.

2.1 Quantitative Evaluation of synthesis results:

Use this script:

python eval_syn_re10k.py [OUTPUT_DIR] (for RealEstate10K)
python eval_syn_mp3d.py [OUTPUT_DIR] (for Matterport3D)

Optional commands:

2.2 Quantitative Evaluation of pose prediction results:

Use this script:

python eval_pose.py [POSE_DIR]

Contact

For any questions about the code or the paper, you can contact zihang.lai at gmail.com.