Unified Line Segment Detection
This repository contains the official PyTorch implementation of the paper: ULSD: Unified Line Segment Detection across Pinhole, Fisheye, and Spherical Cameras .
ULSD is a unified line segment detection method for both distorted and undistorted images from pinhole, fisheye or spherical cameras. With a novel line segment representation based on the Bezier curve, our method can detect arbitrarily distorted line segments. Experimental results on the pinhole, fisheye, and spherical image datasets validate the superiority of the proposed ULSD to the SOTA methods both in accuracy and efficiency.
Method
Wireframe Dataset
YorkUrban Dataset
FPS
sAP5
sAP10
sAP15
msAP
mAPJ
APH
FH
sAP5
sAP10
sAP15
msAP
mAPJ
APH
FH
LSD
8.3
10.8
12.7
10.6
17.2
54.3
61.5
8.5
10.6
12.2
10.4
15.4
49.7
60.0
50.9
DWP
5.8
7.6
8.8
7.4
38.6
65.9
72.2
2.3
3.2
4.1
3.2
23.4
51.6
62.3
2.3
AFM
21.2
26.8
30.2
26.1
24.3
70.1
77.0
8.0
10.3
12.1
10.1
12.5
48.5
63.2
14.3
L-CNN
60.7
64.1
65.6
63.5
59.3
80.3
76.9
25.3
27.2
28.5
27.0
30.3
57.8
61.6
13.7
HAWP
64.5
67.7
69.2
67.1
60.2
83.2
80.2
27.3
29.5
30.8
29.2
31.7
58.8
64.8
30.9
ULSD1 (ours)
65.3
69.0
70.6
68.3
61.6
82.3
80.6
26.6
29.2
30.9
28.9
31.3
56.6
63.4
38.3
ULSD2 (ours)
65.3
69.2
70.9
68.5
61.5
82.5
80.4
27.3
30.2
32.0
29.8
32.3
56.6
63.6
37.2
ULSD3 (ours)
65.0
68.9
70.5
68.1
61.6
82.2
80.1
26.1
28.6
30.4
28.4
31.0
56.1
63.3
37.6
ULSD4 (ours)
65.3
69.2
70.9
68.5
61.4
82.2
80.4
27.7
30.4
32.0
30.0
31.5
56.9
63.8
37.2
Method
F-Wireframe Dataset
F-YorkUrban Dataset
FPS
sAP5
sAP10
sAP15
msAP
mAPJ
sAP5
sAP10
sAP15
msAP
mAPJ
SHT
0.4
0.9
1.3
0.8
2.2
0.4
0.8
1.1
0.8
2.3
0.3
L-CNN
40.0
43.4
45.2
42.9
44.2
18.2
19.9
20.8
19.6
26.5
14.3
HAWP
42.5
46.3
48.0
45.6
43.8
19.5
21.5
22.5
21.2
26.4
31.5
ULSD2 (ours)
59.4
64.3
66.3
63.3
56.4
27.7
30.7
32.4
30.3
33.6
37.2
ULSD3 (ours)
59.7
64.7
66.7
63.7
56.0
27.1
30.2
32.0
29.8
32.9
37.1
ULSD4 (ours)
59.4
64.3
66.3
63.3
56.1
28.8
32.0
33.8
31.5
33.9
36.9
Method
SUN360 Dataset
FPS
sAP5
sAP10
sAP15
msAP
mAPJ
SHT
0.8
1.6
2.3
1.5
4.0
0.2
L-CNN
39.8
42.4
43.6
41.9
41.2
13.4
HAWP
41.6
44.7
45.8
44.0
39.2
25.4
ULSD2 (ours)
63.1
69.4
71.5
68.0
57.8
23.9
ULSD3 (ours)
61.9
69.1
71.1
67.3
57.2
23.7
ULSD4 (ours)
63.8
70.1
71.8
68.6
57.9
23.8
python3
pytorch==1.6.0
CUDA==10.1
opencv, numpy, scipy, matplotlib, argparse, yacs, tqdm, json, multiprocessing, sklearn, tensorboardX
Step-by-step installation
conda create --name ulsd python=3.7
conda activate ulsd
cd <ulsd-path>
git clone https://github.com/lh9171338/Unified-Line-Segment-Detection.git
cd Unified-Line-Segment-Detection
pip install -r requirements.txt
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch
Quickstart with the pretrained model
There are 3 pretrained models (pinhole.pkl , fisheye.pkl , and spherical.pkl ) in Google drive
. Please download them and put in the model/ folder.
There are some testing images in dataset/ folder.
python test.py --config_file pinhole.yaml --dataset_name pinhole --save_image
The result is saved in output/ folder.
Download the json-format dataset.
Convert the dataset from json-format to npz-format.
cd dataset/
python json2npz.py --config_file fisheye.yaml --dataset_name fwireframe --order 2
Generate the ground truth for evaluation.
cd dataset/
python json2npz_gt.py --config_file fisheye.yaml --dataset_name fwireframe
python train.py --config_file fisheye.yaml --dataset_name fwireframe --order 2 [--gpu 0]
python test.py --config_file fisheye.yaml --dataset_name fwireframe --order 2 --model_name best.pkl [--gpu 0] [--save_image] [--evaluate]
Evaluate mAPJ , sAP, and FPS
python test.py --config_file pinhole.yaml --dataset_name wireframe --evaluate
cd metric/
python eval_APH.py --config_file pinhole.yaml --dataset_name wireframe
@misc{li2020ulsd,
title={ULSD: Unified Line Segment Detection across Pinhole, Fisheye, and Spherical Cameras},
author={Hao Li and Huai Yu and Wen Yang and Lei Yu and Sebastian Scherer},
year={2020},
eprint={2011.03174},
archivePrefix={arXiv},
primaryClass={cs.CV}
}