| Vakalopoulou M., Platias C., Papadomanolaki M., Paragios N*., Karantzalos K. | |
NTUA, Greece |
* ECP, France |
| Corresponding authors: mariavak@central.ntua.gr & karank@central.ntua.gr | |
A novel generic graph-based framework was developed addressing simultaneous registration, segmentation and change detection in multisensor data of different spectral, spatial and temporal resolutions. The quite promising experimental results indicated: (i) for the registration task a mean displacement error of less than 2 pixels, (ii) for the segmentation task in most cases completeness and correctness rates above 77% and (iii) for the change detection around or above 70%. The main errors derived from the important relief displacements around the tallest buildings and from the quite similar reflectance spectra of the different man-made objects including ship/vessels. |
Registration Results |
Segmentation Results |
Change Detection Results |
||
The developed framework can take as an input a pair of multisensor images (as the raw unregistered ones below) and perfoms simultaneously a co-registration, co-segmentation and change detection.
Those three problems are adressed concurrently from a single energy formulation (Eq.5), while classification scores are also exploited (Eq.2 & Eq.4) within, based on a deep learning & CNN procedure (classification maps below).
| Raw Unregistered Data |
|
VHR Deimos-2 | March 2015 |
VHR Deimos-2 | May 2015 |
|
| @ 70cm, 4 spectral bands | @ 70cm, 4 spectral bands | |
 |
 |
Google Image Mosaic & Google Map |
Iris Satellite Video Sequence | July 2015 |
|
| @ level 17, RGB | @ 1m, RGB (first frame) | |
 |
 |
|
Deep-learning & CNN |
The developed framework exploits classification scores derived from a deep-learning procedure |
Deimos-2 (May'15) - Classification MapDeep-learning & CNN with 8 classes |
|
 |
 |
 |
|
Deimos-2 (March'15) - Classification MapDeep-learning & CNN with 8 classes |
|
 |
|
 |
|
Iris Video (July'15) - Classification MapDeep-learning & CNN with 7 classes |
|
 |
 |
 |
|