New Modules for Joint Satellite Segmentation/DHM estimation

Takes an 8-band MSI or 3-channel RGB or both as input and produce either DHM estimate or
segmentation estimate (ground, building, tree, water, road) or both jointly. In case DHM is
already known, there is another model provided that takes all 3 modalities (RGB, MSI and DHM)
as input and outputs a segmentation estimate. Refer to the GRSS data page for information on how semantic classes are labelled.

In total, the following 8 new modules are provided:

1. msi_to_agl: Takes MSI as input and produces DHM estimate.
2. msi_to_cls: Takes MSI as input and produces Segmentation estimate.
3. rgb_to_agl: Takes RGB as input and produces DHM estimate.
4. rgb_to_cls: Takes RGB as input and produces Segmentation estimate.
5. rgb_msi_to_agl: Takes RGB and MSI as input and produces DHM estimate.
6. rgb_msi_to_cls: Takes RGB and MSI as input and produces Segmentation estimate.
7. rgb_msi_agl_to_cls: Takes RGB, MSI and DHM as input and produces Segmentation estimate.
8. rgb_msi_to_agl_cls: Takes RGB and MSI as input and jointly estimates DHM and Segmentation labels.

This is all achieved using an end-to-end DCNN based on RBDN, that is trained on the GRSS dataset.

A Dockerized image is also available: docker pull venkai/joint-seg-dhm:latest

Dockerized Modules for Shadow-Removal, Inpainting and MSI2RGB

The current list of modules are as follows:

• Joint Shadow Removal and Shadow Probability Estimation for Satellite Imagery.
• Inpainting
  1. RGB to RGB inpainting
  2. RGBD to RGBD tiled iterative inpainting.
• 8-band MSI (unknown normalization/sensor data) to photo-realistic RGB, shadow-free RGB and shadow probabilities.