Closed dapetri closed 2 years ago
So far implementation is feasible. Removing all pixels whose depth difference is greater than mean removes below 10k pixel.
More improvement potentially in better segmentation. Still large background/surface outliers that result in bended edges in prediction.
So far implementation is feasible. Removing all pixels whose depth difference is greater than mean removes below 10k pixel.
More improvement potentially in better segmentation. Still large background/surface outliers that result in bended edges in prediction.
Pixels in calibrated image set between input and target that have a depth distance bigger than threshold should be removed in postprocessing. if not net smoothes edges between objects