Presentation2019 (Supporting material (Image of sensor pods., Image of the…
Image of sensor pods.
Image of the layout of PSMNet
Image of the layout of CSN
Image of PSMNet plus uncertainty.
Image of the pillar / beam with grids.
Configuration of DNNs
List of uncertainty/confidence measures.
:check: Math of uncertainty for deep learning. (Only the aleatoric uncertainties)
Work with downsampled images?
:check:Make sure the image size does not affacts LIDAR points gathering.
Visualize the LIDAR point cloud.
:check:Make it work with downsampled images.
:check:A table to record different trainnings.
Training scripts. The filenames are not good.
Deep Learning with Uncertainty
Record and plot the sigma values
:checkered_flag:The loss may become negative. Is this the desired behavior? Kind of.
:red_cross:May use larger learing rate. Training exploded with lr=0.01
:checkered_flag:Find out the unit/value range of sigma. Seems to be the unit of pixel
Compare with different methods
Prepare the data
Open available / benchmark data
Findout the statistics of the data
201812 bridge scan
NSH wall scan
201810 Shinizu beam
OpenCV SGM for single pair of images.
:warning:Determine the image size for different methods!
:checkered_flag:512x1024 for now
PSMNet with uncertainty
Need more better way to visualize.
:check:Camera matrix and Q matrix. How to read these data into different methods.
Plane / Patch matching methods
:check:Limit minimum disparity value. (Limit lower bound may not be a good idea if we know we wil see objectives at long distance.)
Output point cloud in PLY format.
:check:Read camera matrix.
:check:Add search range for individual pixel
Deep Learning methods
Without 3D cost volume
With 3D cost volume
Bilateral window matcher
Only have the results of matching in single line.
Learn confidence? :warning:Following ups of the discussion with Basti.