STCrowd
STCrowd copied to clipboard
Questions about implementation of DHA details
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.
- 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.
- 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.