Pairwise depth estimation
PDE pipeline, two jetson devices in a master slave configuration. Input is image data (video or singular). Output is image data, stream or frame + Depth map, stream or frame. The PDE pipeline consists of the following stages
- Depth map (aka depthmap), 2D matrix of floating point depth values.
- Master / Slave pair. The PDE pipleine is designed as a pair of jetsons. One being designated master, the other is designated slave. The Master initiates the protocol, ths slave starts by listening for a connection by the master.
- Image rectification, the process of reprojecting the master's and slave's image onto a common plane.
- Image de-rectification, the inverse process of rectification.
- Image crosssharing, the process of master (1) transmitting the first half (+ extra padding) of its image data to the slave. The slave (2) sends its second half of its image to the master.
- Depth estimation, input two half images, output two half depth maps. One for each camera view. This is currently only planned as a stub. Convert to grayscale.
- Depth map crosssharing. Master sends to slave the estimate half depthmap corresponding to the slaves camera view. Slave sends to master the estimate half depthmap corresponding to the masters view.
Depth map blending. On each device, the two half depthmaps corresponding to the devices camera view compose into the complete depth map.
- PDE output. An gstreamer appsink implementation particular for its purpose.
- Plain screen device. Output: 2 video streams, image + depth. Client uses depth to select / manipulate sections of the image for visualization. Meta contained within the streams is restricted to timestamps.
- Dimenco. The big auto-steroscopic display. Output: 2 video streams, image + depth (same view). The client uses a library (ver:/rogers_code) to construct a special composite image to send to the display.
- Steroscopic displays (2 render views). Output: A selection of N x 2 video streams (image + depth). Each stream carries metadata (camera geometry matrices and timing info). N views. The client uses camera info to select streams and render the displayed views. The process of stream selection is still unspecified.
$Id: pde.vocabulary.md 651 2018-03-15 15:30:15Z elidim $