Some Questions About Implementation Details.

[wbhwbh]

Thank you for every incredible work you have contributed to the VLN community. I am really interested in this project and have some questions about the implementation details as it is not mentioned in the paper.

• What's the data type of the semantic map (referred as $M$ in Figure 1 in the paper) which is input into the convolutional network? Is it a RGB image or a numpy array that each element represents for a category?
• As not mentioned in the paper Semantic-MapNet, how are the trajectory and open space semantics in $M$ obtained?
• The map is also powered with semantic information (denoted at the bottom right) as you wirte in the caption of Figure 1. Do you explicitly tell the model these information or let model learn these semantic information by itself?