Pytorch FCGF_spconv

This repo contains an unofficial implementation for Fully Convolutional Geometric Features (FCGF, ICCV 19) based on spconv v2.1.22 instead of MinkowskiEngine. The original observation is the SLOW training speed during runing the original FCGF repo. As another alternative for sparse convolution, spconv has been widely used for tasks such as 3D object detection, etc. Here we follow the network paradigm of FCGF but build the model using convolutional layers provided by spconv.

Requirements

# pytorch
plyfile
nibabel
easydict
open3d==0.10.0
scipy
tensorboardx
tqdm

install above packages by:

pip install -r requirements.txt

Also, install spconv via pip (CUDA11.1):

pip install spconv-cu111

Data Preparation

Download the 3DMatch provided by FCGF and organize them as follows:

├── FCGF_data  
│   ├──	threedmatch  
        ├── 7-scenes-redkitchen@seq-01_000.npz
        ├── 7-scenes-redkitchen@seq-01_001.npz
        └── ... 
│   ├── threedmatch_test/
        ├── 7-scenes-redkitchen
        ├── 7-scenes-redkitchen-evaluation
        └── ...

Training on 3DMatch

sh scripts/train_3dmatch.sh

Testing on 3DMatch

sh scripts/evaluate.sh

Performance

The total training time consumes about:

Methods	Training Time	Pretrained Weight	Epochs
FCGF	~41.6h (25min per epoch)	2019-08-19_06-17-41.pth	100
FCGF_spconv	~11.5 h (23min per epoch)	ckpt_30.pth	30

We report Inlier Ratio, Feature Matching Recall and Registration Recall as three main metrics to compare the FCGF_spconv with the original FCGF:

Methods	Voxel Size	sample	Mutual Selection	Inlier Ratio	Feature Matching Recall	Registration Recall
FCGF (paper)	0.025			~	0.952	0.82
FCGF	0.025	5k	❌	0.341	0.956	0.8343
FCGF_spconv	0.025	5k	❌	0.2889	0.928	0.8757
FCGF_spconv (on 3DMatchRotated)	0.25	5k	❌	0.0924	0.618	0.6739

Generalization Ability (From 3DMatch -> ETH)

We test the generalization ability of FCGF_spconv on the challenging outdoor dataset ETH, which is acquired by static terrestrial scanners and dominated by outdoor vegetation, such as trees and bushes. Here we compute the Feature Matching Recall (mutual) to show the generalization ability.

Download the ETH dataset processed by YOHO, then run:

python benchmark_ETH.py

Method	Gazebo_summer	Gazebo_winter	Wood_autumn	Wood_summer	Average
FCGF	22.8	10.0	14.8	16.8	16.1
D3Feat	85.9	63.0	49.6	48.0	56.3
SpinNet	92.9	91.7	92.2	94.4	92.8
DIP	90.8	88.6	96.5	95.2	92.8
FCGF_spconv	33.7	17.3	21.7	28.0	25.2

There is a HUGE gap between the patched-based and fully-convolutional based networks with respect to the generalization ability :(

Registration Demo

modify the src_path, tgt_path, voxel_size as well as n_sample in demo.py to register your custom point cloud data.

python demo.py

Reference

FCGF, spconv, DIP.

qsisi / FCGF_spconv