Qinghui Liu, Michael Kampffmeyer,
Robert Jenssen, and Arnt-Børre Salberg
July. 29, 2019
Motivation
Airborne laser scanning (ALS) Data
3D point cloud ALS data
Digital elevation model (DEM)
DEM created from 3D point cloud ALS data
DEM resolution: 0.5m x 0.5m
Gradient image and gound truth
Road-3
Road-1
Road-2
Gradient (slope) image of the DEM
Ground truth (binary and 3-classes)
Example of the DDCM architecture
composed of N {1, 2, 3, ..., N} DC blocks with various dilation rates.
Dense Dilated Convolutions Merging (DDCM) module
Dense Dilated Convolutions Merging (DDCM) module
1-D example of DDCM [1, 2, 4]
DDCM-R50 model with pretrained ResNet50
Single Task DDCM-Net for Land Cover Mapping.
Joint-Task DDCM-Net for Road Mapping
JT-DDCM-R50 model with 2 more joint learning tasks.
Iterative-Random Weighting Losses
We sample the weights
and
uniformly between
denotes the current iteration of training,
the MFB weighted cross entropy loss for the main task
the binary cross entropy loass for the classification task
lovasz-softmax loss for the binary segmentation task
Where,
: the median frequency balancing (MFB) weights.
[13] M. B. A. R. T. Matthew and B. Blaschko, “The Lovasz-softmax loss: ´A tractable surrogate for the optimization of the intersection-over-union measure in neural networks,” 2018.
ISPRS Vaihingen: http://www2.isprs.org/commissions/comm3/wg4/2d-sem-label-vaihingen.html
ISPRS Potsdam: http://www2.isprs.org/commissions/comm3/wg4/2d-sem-label-potsdam.html
ISPRS Vaihingen
ALS (Gradient + Hillshade)
Datasets
Comparison: Parameters Size, GFLOPs and mIoU
(-1.9%)
(+9x)
(+4x)
(GFLOPs measured on input size 3 x 256 x256),
Computational Cost
Vaihingen hold-out (17 unseen images) test results.
Performance on ISPRS dataset
(+4.6%)
(+5.6%)
Test Results on Our Lidar dataset
Test image
Full-class GT
DDCM-R50
JT-DDCM-R50
Analysis of Results
Conclusions and Future Work
Conlusions
Future work
Thanks for your attention!
Q & A
Brian.Liu@nr.no