Code release for Local Light Field Fusion at SIGGRAPH 2019

Overview





Local Light Field Fusion

Project | Video | Paper

Tensorflow implementation for novel view synthesis from sparse input images.

Local Light Field Fusion: Practical View Synthesis with Prescriptive Sampling Guidelines
Ben Mildenhall*1, Pratul Srinivasan*1, Rodrigo Ortiz-Cayon2, Nima Khademi Kalantari3, Ravi Ramamoorthi4, Ren Ng1, Abhishek Kar2
1UC Berkeley, 2Fyusion Inc, 3Texas A&M, 4UC San Diego
*denotes equal contribution
In SIGGRAPH 2019

Table of Contents

Installation TL;DR: Setup and render a demo scene

First install docker (instructions) and nvidia-docker (instructions).

Run this in the base directory to download a pretrained checkpoint, download a Docker image, and run code to generate MPIs and a rendered output video on an example input dataset:

bash download_data.sh
sudo docker pull bmild/tf_colmap
sudo docker tag bmild/tf_colmap tf_colmap
sudo nvidia-docker run --rm --volume /:/host --workdir /host$PWD tf_colmap bash demo.sh

A video like this should be output to data/testscene/outputs/test_vid.mp4:

If this works, then you are ready to start processing your own images! Run

sudo nvidia-docker run -it --rm --volume /:/host --workdir /host$PWD tf_colmap

to enter a shell inside the Docker container, and skip ahead to the section on using your own input images for view synthesis.

Full Installation Details

You can either install the prerequisites by hand or use our provided Dockerfile to make a docker image.

In either case, start by downloading this repository, then running the download_data.sh script to download a pretrained model and example input dataset:

bash download_data.sh

After installing dependencies, try running bash demo.sh from the base directory. (If using Docker, run this inside the container.) This should generate the video shown in the Installation TL;DR section at data/testscene/outputs/test_vid.mp4.

Manual installation

  • Install CUDA, Tensorflow, COLMAP, ffmpeg
  • Install the required Python packages:
pip install -r requirements.txt
  • Optional: run make in cuda_renderer/ directory.
  • Optional: run make in opengl_viewer/ directory. You may need to install GLFW or some other OpenGL libraries. For GLFW:
sudo apt-get install libglfw3-dev

Docker installation

To build the docker image on your own machine, which may take 15-30 mins:

sudo docker build -t tf_colmap:latest .

To download the image (~6GB) instead:

sudo docker pull bmild/tf_colmap
sudo docker tag bmild/tf_colmap tf_colmap

Afterwards, you can launch an interactive shell inside the container:

sudo nvidia-docker run -it --rm --volume /:/host --workdir /host$PWD tf_colmap

From this shell, all the code in the repo should work (except opengl_viewer).

To run any single command <command...> inside the docker container:

sudo nvidia-docker run --rm --volume /:/host --workdir /host$PWD tf_colmap <command...>

Using your own input images for view synthesis

Our method takes in a set of images of a static scene, promotes each image to a local layered representation (MPI), and blends local light fields rendered from these MPIs to render novel views. Please see our paper for more details.

As a rule of thumb, you should use images where the maximum disparity between views is no more than about 64 pixels (watch the closest thing to the camera and don't let it move more than ~1/8 the horizontal field of view between images). Our datasets usually consist of 20-30 images captured handheld in a rough grid pattern.

Quickstart: rendering a video from a zip file of your images

You can quickly render novel view frames and a .mp4 video from a zip file of your captured input images with the zip2mpis.sh bash script.

bash zip2mpis.sh <zipfile> <your_outdir> [--height HEIGHT]

height is the output height in pixels. We recommend using a height of 360 pixels for generating results quickly.

General step-by-step usage

Begin by creating a base scene directory (e.g., scenedir/), and copying your images into a subdirectory called images/ (e.g., scenedir/images).

1. Recover camera poses

This script calls COLMAP to run structure from motion to get 6-DoF camera poses and near/far depth bounds for the scene.

python imgs2poses.py <your_scenedir>

2. Generate MPIs

This script uses our pretrained Tensorflow graph (make sure it exists in checkpoints/papermodel) to generate MPIs from the posed images. They will be saved in <your_mpidir>, a directory will be created by the script.

python imgs2mpis.py <your_scenedir> <your_mpidir> \
    [--checkpoint CHECKPOINT] \
    [--factor FACTOR] [--width WIDTH] [--height HEIGHT] [--numplanes NUMPLANES] \
    [--disps] [--psvs] 

You should set at most one of factor, width, or height to determine the output MPI resolution (factor will scale the input image size down an integer factor, eg. 2, 4, 8, and height/width directly scale the input images to have the specified height or width). numplanes is 32 by default. checkpoint is set to the downloaded checkpoint by default.

Example usage:

python imgs2mpis.py scenedir scenedir/mpis --height 360

3. Render novel views

You can either generate a list of novel view camera poses and render out a video, or you can load the saved MPIs in our interactive OpenGL viewer.

Generate poses for new view path

First, generate a smooth new view path by calling

python imgs2renderpath.py <your_scenedir> <your_posefile> \
	[--x_axis] [--y_axis] [--z_axis] [--circle][--spiral]

<your_posefile> is the path of an output .txt file that will be created by the script, and will contain camera poses for the rendered novel views. The five optional arguments specify the trajectory of the camera. The xyz-axis options are straight lines along each camera axis respectively, "circle" is a circle in the camera plane, and "spiral" is a circle combined with movement along the z-axis.

Example usage:

python imgs2renderpath.py scenedir scenedir/spiral_path.txt --spiral

See llff/math/pose_math.py for the code that generates these path trajectories.

Render video with CUDA

You can build this in the cuda_renderer/ directory by calling make.

Uses CUDA to render out a video. Specify the height of the output video in pixels (-1 for same resolution as the MPIs), the factor for cropping the edges of the video (default is 1.0 for no cropping), and the compression quality (crf) for the saved MP4 file (default is 18, lossless is 0, reasonable is 12-28).

./cuda_renderer mpidir <your_posefile> <your_videofile> height crop crf

<your_videofile> is the path to the video file that will be written by FFMPEG.

Example usage:

./cuda_renderer scenedir/mpis scenedir/spiral_path.txt scenedir/spiral_render.mp4 -1 0.8 18

Render video with Tensorflow

Use Tensorflow to render out a video (~100x slower than CUDA renderer). Optionally, specify how many MPIs are blended for each rendered output (default is 5) and what factor to crop the edges of the video (default is 1.0 for no cropping).

python mpis2video.py <your_mpidir> <your_posefile> videofile [--use_N USE_N] [--crop_factor CROP_FACTOR]

Example usage:

python mpis2video.py scenedir/mpis scenedir/spiral_path.txt scenedir/spiral_render.mp4 --crop_factor 0.8

Interactive OpenGL viewer

Controls:

  • ESC to quit
  • Move mouse to translate in camera plane
  • Click and drag to rotate camera
  • Scroll to change focal length (zoom)
  • 'L' to animate circle render path

The OpenGL viewer cannot be used in the Docker container.

You need OpenGL installed, particularly GLFW:

sudo apt-get install libglfw3-dev

You can build the viewer in the opengl_viewer/ directory by calling make.

General usage (in opengl_viewer/ directory) is

./opengl_viewer mpidir

Using your own poses without running COLMAP

Here we explain the poses_bounds.npy file format. This file stores a numpy array of size Nx17 (where N is the number of input images). You can see how it is loaded in the three lines here. Each row of length 17 gets reshaped into a 3x5 pose matrix and 2 depth values that bound the closest and farthest scene content from that point of view.

The pose matrix is a 3x4 camera-to-world affine transform concatenated with a 3x1 column [image height, image width, focal length] to represent the intrinsics (we assume the principal point is centered and that the focal length is the same for both x and y).

The right-handed coordinate system of the the rotation (first 3x3 block in the camera-to-world transform) is as follows: from the point of view of the camera, the three axes are [down, right, backwards] which some people might consider to be [-y,x,z], where the camera is looking along -z. (The more conventional frame [x,y,z] is [right, up, backwards]. The COLMAP frame is [right, down, forwards] or [x,-y,-z].)

If you have a set of 3x4 cam-to-world poses for your images plus focal lengths and close/far depth bounds, the steps to recreate poses_bounds.npy are:

  1. Make sure your poses are in camera-to-world format, not world-to-camera.
  2. Make sure your rotation matrices have the columns in the correct coordinate frame [down, right, backwards].
  3. Concatenate each pose with the [height, width, focal] intrinsics vector to get a 3x5 matrix.
  4. Flatten each of those into 15 elements and concatenate the close and far depths.
  5. Stack the 17-d vectors to get a Nx17 matrix and use np.save to store it as poses_bounds.npy in the scene's base directory (same level containing the images/ directory).

This should explain the pose processing after COLMAP.

Troubleshooting

  • PyramidCU::GenerateFeatureList: an illegal memory access was encountered: Some machine configurations might run into problems running the script imgs2poses.py. A solution to that would be to set the environment variable CUDA_VISIBLE_DEVICES. If the issue persists, try uncommenting this line to stop COLMAP from using the GPU to extract image features.
  • Black screen: In the latest versions of MacOS, OpenGL initializes a context with a black screen until the window is dragged or resized. If you run into this problem, please drag the window to another position.
  • COLMAP fails: If you see "Could not register, trying another image", you will probably have to try changing COLMAP optimization parameters or capturing more images of your scene. See here.

Citation

If you find this useful for your research, please cite the following paper.

@article{mildenhall2019llff,
  title={Local Light Field Fusion: Practical View Synthesis with Prescriptive Sampling Guidelines},
  author={Ben Mildenhall and Pratul P. Srinivasan and Rodrigo Ortiz-Cayon and Nima Khademi Kalantari and Ravi Ramamoorthi and Ren Ng and Abhishek Kar},
  journal={ACM Transactions on Graphics (TOG)},
  year={2019},
}
Scenic: A Jax Library for Computer Vision and Beyond

Scenic Scenic is a codebase with a focus on research around attention-based models for computer vision. Scenic has been successfully used to develop c

Google Research 1.6k Dec 27, 2022
This repository is for Contrastive Embedding Distribution Refinement and Entropy-Aware Attention Network (CEDR)

CEDR This repository is for Contrastive Embedding Distribution Refinement and Entropy-Aware Attention Network (CEDR) introduced in the following paper

phoenix 3 Feb 27, 2022
PyTorch code for our paper "Gated Multiple Feedback Network for Image Super-Resolution" (BMVC2019)

Gated Multiple Feedback Network for Image Super-Resolution This repository contains the PyTorch implementation for the proposed GMFN [arXiv]. The fram

Qilei Li 66 Nov 03, 2022
Differential Privacy for Heterogeneous Federated Learning : Utility & Privacy tradeoffs

Differential Privacy for Heterogeneous Federated Learning : Utility & Privacy tradeoffs In this work, we propose an algorithm DP-SCAFFOLD(-warm), whic

19 Nov 10, 2022
Progressive Coordinate Transforms for Monocular 3D Object Detection

Progressive Coordinate Transforms for Monocular 3D Object Detection This repository is the official implementation of PCT. Introduction In this paper,

58 Nov 06, 2022
A Review of Deep Learning Techniques for Markerless Human Motion on Synthetic Datasets

HOW TO USE THIS PROJECT A Review of Deep Learning Techniques for Markerless Human Motion on Synthetic Datasets Based on DeepLabCut toolbox, we run wit

1 Jan 10, 2022
A repository for benchmarking neural vocoders by their quality and speed.

License The majority of VocBench is licensed under CC-BY-NC, however portions of the project are available under separate license terms: Wavenet, Para

Meta Research 177 Dec 12, 2022
xitorch: differentiable scientific computing library

xitorch is a PyTorch-based library of differentiable functions and functionals that can be widely used in scientific computing applications as well as deep learning.

24 Apr 15, 2021
High performance Cross-platform Inference-engine, you could run Anakin on x86-cpu,arm, nv-gpu, amd-gpu,bitmain and cambricon devices.

Anakin2.0 Welcome to the Anakin GitHub. Anakin is a cross-platform, high-performance inference engine, which is originally developed by Baidu engineer

514 Dec 28, 2022
HiFT: Hierarchical Feature Transformer for Aerial Tracking (ICCV2021)

HiFT: Hierarchical Feature Transformer for Aerial Tracking Ziang Cao, Changhong Fu, Junjie Ye, Bowen Li, and Yiming Li Our paper is Accepted by ICCV 2

Intelligent Vision for Robotics in Complex Environment 55 Nov 23, 2022
https://sites.google.com/cornell.edu/recsys2021tutorial

Counterfactual Learning and Evaluation for Recommender Systems (RecSys'21 Tutorial) Materials for "Counterfactual Learning and Evaluation for Recommen

yuta-saito 45 Nov 10, 2022
Food Drinks and groceries Images Multi Lingual (FooDI-ML) dataset.

Food Drinks and groceries Images Multi Lingual (FooDI-ML) dataset.

41 Jan 04, 2023
SSD: A Unified Framework for Self-Supervised Outlier Detection [ICLR 2021]

SSD: A Unified Framework for Self-Supervised Outlier Detection [ICLR 2021] Pdf: https://openreview.net/forum?id=v5gjXpmR8J Code for our ICLR 2021 pape

Princeton INSPIRE Research Group 113 Nov 27, 2022
The NEOSSat is a dual-mission microsatellite designed to detect potentially hazardous Earth-orbit-crossing asteroids and track objects that reside in deep space

The NEOSSat is a dual-mission microsatellite designed to detect potentially hazardous Earth-orbit-crossing asteroids and track objects that reside in deep space

John Salib 2 Jan 30, 2022
UpChecker is a simple opensource project to host it fast on your server and check is server up, view statistic, get messages if it is down. UpChecker - just run file and use project easy

UpChecker UpChecker is a simple opensource project to host it fast on your server and check is server up, view statistic, get messages if it is down.

Yan 4 Apr 07, 2022
Multi-Anchor Active Domain Adaptation for Semantic Segmentation (ICCV 2021 Oral)

Multi-Anchor Active Domain Adaptation for Semantic Segmentation Munan Ning*, Donghuan Lu*, Dong Wei†, Cheng Bian, Chenglang Yuan, Shuang Yu, Kai Ma, Y

Munan Ning 36 Dec 07, 2022
Unofficial implementation of Fast-SCNN: Fast Semantic Segmentation Network

Fast-SCNN: Fast Semantic Segmentation Network Unofficial implementation of the model architecture of Fast-SCNN. Real-time Semantic Segmentation and mo

Philip Popien 69 Aug 11, 2022
REBEL: Relation Extraction By End-to-end Language generation

REBEL: Relation Extraction By End-to-end Language generation This is the repository for the Findings of EMNLP 2021 paper REBEL: Relation Extraction By

Babelscape 222 Jan 06, 2023
AQP is a modular pipeline built to enable the comparison and testing of different quality metric configurations.

Audio Quality Platform - AQP An Open Modular Python Platform for Objective Speech and Audio Quality Metrics AQP is a highly modular pipeline designed

Jack Geraghty 24 Oct 01, 2022
This project is based on RIFE and aims to make RIFE more practical for users by adding various features and design new models

CPM 项目描述 CPM(Chinese Pretrained Models)模型是北京智源人工智能研究院和清华大学发布的中文大规模预训练模型。官方发布了三种规模的模型,参数量分别为109M、334M、2.6B,用户需申请与通过审核,方可下载。 由于原项目需要考虑大模型的训练和使用,需要安装较为复杂

hzwer 190 Jan 08, 2023