Fully Convolutional Geometric Features (FCGF)

Pretrained weights for Fully Convolutional Geometric Features, ICCV 2019.

• Paper: Fully Convolutional Geometric Features
• Code: chrischoy/FCGF

FCGF extracts dense 3D geometric features in a single forward pass through a 3D fully-convolutional network, built on MinkowskiEngine. These checkpoints are the official weights released with the paper, previously hosted on node1.chrischoy.org and mirrored here.

Model Zoo

All models use the ResUNetBN2C backbone.

File	Normalized Feature	Dataset	Voxel Size	Feature Dimension	Performance
2019-08-19_06-17-41.pth	True	3DMatch	2.5cm (0.025)	32	FMR: 0.9578 +- 0.0272
2019-09-18_14-15-59.pth	True	3DMatch	2.5cm (0.025)	16	FMR: 0.9442 +- 0.0345
2019-08-16_19-21-47.pth	True	3DMatch	5cm (0.05)	32	FMR: 0.9372 +- 0.0332
2019-07-31_19-30-19.pth	False	KITTI	20cm (0.2)	32	RTE: 0.0534m, RRE: 0.1704°
2019-07-31_19-37-00.pth	False	KITTI	30cm (0.3)	32	RTE: 0.0607m, RRE: 0.2280°
KITTI-v0.3-ResUNetBN2C-conv1-5-nout16.pth	True	KITTI	30cm (0.3)	16	RTE: 0.0670m, RRE: 0.2295°
KITTI-v0.3-ResUNetBN2C-conv1-5-nout32.pth	True	KITTI	30cm (0.3)	32	RTE: 0.0639m, RRE: 0.2253°

redkitchen-20.ply is the sample point cloud used by demo.py in the FCGF repo.

Usage

import torch
from urllib.request import urlretrieve

urlretrieve(
    "https://huggingface.co/chrischoy/FCGF/resolve/main/2019-09-18_14-15-59.pth",
    "ResUNetBN2C-16feat-3conv.pth")

checkpoint = torch.load("ResUNetBN2C-16feat-3conv.pth")

See chrischoy/FCGF for the full model definition (ResUNetBN2C) and feature extraction code.

Citation

@inproceedings{FCGF2019,
    author = {Christopher Choy and Jaesik Park and Vladlen Koltun},
    title = {Fully Convolutional Geometric Features},
    booktitle = {ICCV},
    year = {2019},
}