1

I'm training a CNN with images which have lots of horizontal black lines (due to the nature of the sensor). I'm thinking in removing this artifacts by some kind of preprocessing (interpolation, median filters...). The thing is: does it make sense, since the CNN tries to apply optimal filtering? (if some 2D filtering is intended to be done just before the CNN, is just adding a deterministic layer at the beginning of the net...)

ignatius
  • 1,638
  • 7
  • 21
  • Do the black lines actually tell you anything? Is there any signal to extract from them, or are they essentially just _noise_? – n1k31t4 May 08 '18 at 09:54
  • They are artifacts/noise: the images come from 2D projections of a lidar sensor, where the vertical dimension of the image represents the elevation angle of the 3D point cloud. Since the LiDAR sensor has no continuity in the elevation range, there are some black lines. I know there would be other approaches, like just removing these black lines or controlling them in the 2D projection stage, but if it is not desired to do such things, does it make sense pre-filter the images? – ignatius May 08 '18 at 10:59
  • Talking about LiDAR data, the spacing between these horizontal lines does usually contain some information; spacings are proportional to the angles between your individual lasers. 2D projections of pointclouds will sometimes be very sparse, depending on your point density. I would recommend you first make certain you really cannot encode this information in the images somehow (e.g. using colours to represent elevation), before removing it. Otherwise, just try it out with/without pre-processing, using comparable model architectures of course - then compare results. – n1k31t4 May 08 '18 at 11:37
  • I have the results for the non-processed images. I'm voting for doing a linear interpolation over the vertical direction and asap I get the results I'll post them – ignatius May 08 '18 at 12:57

1 Answers1

0

That is an empirical question which can be answered through cross validation.

Consider keeping or removing sensor artifacts a hyperparameter. See if the different values of that hyperparameter impact performance on an evaluation metric on a hold-out dataset.

Brian Spiering
  • 20,142
  • 2
  • 25
  • 102