preview.mp4
This is the official implementation for FlowMap: High-Quality Camera Poses, Intrinsics, and Depth via Gradient Descent by Cameron Smith*, David Charatan*, Ayush Tewari, and Vincent Sitzmann.
Check out the project website here.
To get started on Linux, create a Python virtual environment:
python3.11 -m venv venv
source venv/bin/activate
pip install -r requirements.txtFor pretraining, make sure GMFlow is installed as a submodule:
git submodule update --init --recursive
If the above requirements don't work, you can try requirements_exact.txt instead.
The main entry point is flowmap/overfit.py. Call it via:
python3 -m flowmap.overfit dataset=images dataset.images.root=path/to/folder/with/imagesMake sure the virtual environment has been activated via source venv/bin/activate first.
The checkpoint we used to initialize FlowMap can be found here. To train your own, download the Real Estate 10k and CO3Dv2 datasets and run the following script:
python3 -m flowmap.pretrainSome of the videos in the Real Estate 10k dataset are no longer publicly available. Reach out to us via email if you want our downloaded version of the dataset.
We evaluated FlowMap using video subsets of the Local Light Field Fusion (LLFF), Mip-NeRF 360, and Tanks & Temples datasets. We've uploaded a compilation of these datasets here.
Dataset Details
These are the LLFF scenes from the NeRF paper, which were originally uploaded here. We used all 8 scenes (fern, flower, fortress, horns, leaves, orchids, room, and trex).
These are scenes from the Mip-NeRF 360 paper, which were originally uploaded here. We used the bonsai, counter, and kitchen scenes. The original kitchen scene consists of several concatenated video sequences; for FlowMap, we use the first one (65 frames). We also included the garden scene, which is somewhat video-like, but contain large jumps that make optical flow estimation struggle.
We used all scenes from the Tanks & Temples dataset: auditorium, ballroom, barn, caterpillar, church, courthouse, family, francis, horse, ignatius, lighthouse, m60, meetingroom, museum, palace, panther, playground, temple, train, and truck. We preprocessed the raw videos from the dataset using the script at flowmap/subsample.py. This script samples 150 frames from the first minute of video evenly based on mean optical flow.
Each ablation shown in the paper has a Hydra configuration at config/experiment. For example, to run the ablation where point tracking is disabled, add +experiment=ablation_no_tracks to the overfitting command. Note that you can stack most of the ablations, e.g., +experiment=[ablation_no_tracks,ablation_random_initialization].
We used a modified version of the original 3D Gaussian Splatting code that backpropagates into camera positions in order to generate the novel view synthesis results shown in the paper. You can find it here.
Some of the code used to generate the tables and figures in the paper can be found in the assets folder. We used this code alongside Figma and LaTeXiT to create the figures in the paper. You can find our Figma file here. See .vscode/launch.json for the commands needed to run figure generation.
@inproceedings{smith24flowmap,
title={FlowMap: High-Quality Camera Poses, Intrinsics, and Depth via Gradient Descent},
author={Cameron Smith and David Charatan and Ayush Tewari and Vincent Sitzmann},
year={2024},
booktitle={arXiv},
}This work was supported by the National Science Foundation under Grant No. 2211259, by the Singapore DSTA under DST00OECI20300823 (New Representations for Vision), by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) under 140D0423C0075, by the Amazon Science Hub, and by IBM. The Toyota Research Institute also partially supported this work. The views and conclusions contained herein reflect the opinions and conclusions of its authors and no other entity.
