User's Manual

Post Processing
User Guide 26
5 Post Processing
Overview
Librealsense implementation includes post-processing filters to enhance the quality of
depth data and reduce noise levels. All the filters are implemented in the library core
as independent blocks to be used in the customer code
Design and Implementation
Post-processing modules are encapsulated into self-contained processing blocks, which
provide for the following key requirements:
1. Synchronous/Asynchronous invocation
2. Internal frames memory/lifetime management
The filters are capable to receive frames from different sources, though performance-
wise, maintaining filter pipe per camera source is recommended. Internally, the filters
preserve the original data and generate a new (filtered) frame to pass on. The re-
generation of new frames allows to share frame among different consumers (thread)
without the risk of the data being overwritten by another user.
All filters support discreet as well as floating point input data formats. The floating point
inputs are utilized by D400 stereo-based Depth cameras that support Disparity data
representation. The discreet version of the filters is primarily used with SR300 camera,
but also can be applied to D400 devices (though not recommended).
Filters Description
5.3.1 Decimation Filter
Effectively reduces the depth scene complexity. The filter operates on patches of [2x2]
to [8x8] pixels using median filter as pass-criteria. The image size is scaled down
proportionally in both dimensions to preserve the original aspect ratio.
Internally, the filter imposes 4-pixel block alignment for the output frame size width
and height. E.g. for input size (1280X720) and scale factor 3 the output size calculation
is:
[1280,720]/3 -> [426.6666667, 240] -> [428,240]
*The padded rows/columns are zero-filled.
After the resulted frame is produced, the frame intrinsic parameters are recalculated to
compensate for the resolution changes.
The filter also provides some hole filling capability, as the median filter uses valid (non-
zero) pixels only.