Questions about implementation of DHA details

[nghiabka]

Thank for excellent work. This is suitable for pedestrian detection with occlusion.
I am re-implementing DHA method that comnbine to CenterPoint. I have some questions.

1. Spatial attention moudle applies for regular heatmap as input or feature map output of CenterPoint as input.
   2.What is method used for upsampling and down sampling? I am considering between two methods

• non-learnable method: max-pooling for down-sampling, interpolation for upsampling.
• learable method: covolution for downsampling, convolution transpose for upsampling.

3. what is value of FEATURE_MAP_STRIDE in TARGET_ASSIGNER_CONFIG ? 8 or 4 or 2

I am very appreciative to receive suggestions from you.