Having multiple types of sensors in a common rig can introduce challenges based on sampling frequencies, which prevents the perfect one-to-one frame matching of the parallel information streams. Interpolation in the time domain can ease the association between the samples of the modalities. With its slower refresh rate compared to cameras and accurate 3D distance measurements, LIDAR can be a potential target for resampling in time. One of its drawbacks is that it provides output in an unstructured format, which needs additional consideration. To handle the time-domain interpolation of noncontinuous, varying-sized geometry, we voted for the structured spherical representation of the LIDAR points on which we were utilizing a convolutional neural network architecture. Here we show that the geometry can be preserved during interpolation by a modified autoencoder just as shapes are handled in the original image-processing use case.