Splatting SA: Direct Rendering of Synthetic Aperture Imagery
Om publikasjonen
Synthetic aperture radar (SAR) and sonar (SAS)
imaging based on conventional signal processing techniques
are now mature technologies capable of producing high-quality
images. When combined with 2-D and 3-D rendering techniques,
they can yield highly engaging visualizations. In the field of
optical 3-D reconstruction, direct rendering approaches have
been popularized by methods, such as neural radiance fields
and Gaussian splatting. Inspired by this, we propose a novel
direct rendering approach to synthetic aperture (SA) imaging
that performs conventional SA signal accumulation directly in
camera space. In our method, a camera observer is defined, and
ray marching is employed to determine the imaging coordinates.
A signal primitive representation is introduced, and each signal
is coherently delayed and summed in a highly parallelizable
manner. The output is a raster image render generated from the
perspective defined by the camera model, with full SA processing
gain but without the explicit beamforming and intermediate
processing steps characteristic of conventional SA techniques.
Our proposed method is a proof of concept applicable to a range
of use cases, including interactive data inspection with dynamic
processing adjustments and real-time visualization of incoming
data streams. We demonstrate its effectiveness on real SA sonar
and radar datasets.
imaging based on conventional signal processing techniques
are now mature technologies capable of producing high-quality
images. When combined with 2-D and 3-D rendering techniques,
they can yield highly engaging visualizations. In the field of
optical 3-D reconstruction, direct rendering approaches have
been popularized by methods, such as neural radiance fields
and Gaussian splatting. Inspired by this, we propose a novel
direct rendering approach to synthetic aperture (SA) imaging
that performs conventional SA signal accumulation directly in
camera space. In our method, a camera observer is defined, and
ray marching is employed to determine the imaging coordinates.
A signal primitive representation is introduced, and each signal
is coherently delayed and summed in a highly parallelizable
manner. The output is a raster image render generated from the
perspective defined by the camera model, with full SA processing
gain but without the explicit beamforming and intermediate
processing steps characteristic of conventional SA techniques.
Our proposed method is a proof of concept applicable to a range
of use cases, including interactive data inspection with dynamic
processing adjustments and real-time visualization of incoming
data streams. We demonstrate its effectiveness on real SA sonar
and radar datasets.