Hyperlayer visualization

[yihuaweiyihua]

Thank you very much for your outstanding work!
However, I have two questions regarding the visualization of super layers. Firstly, according to the default configuration of two super layers and eight heat maps, were the three visualization maps selected in the paper the most representative and explanatory ones.
Secondly, the heat map of the super layer is one sixteenth of the original image. How did you enlarge it and overlay it onto the original image.

[ahmedelbamby-aast]

1. About the 3 visualizations in the paper (from default 2 super layers / 8 heat maps):

• In practice, those 3 maps are usually selected as representative and explanatory examples, not the only valid maps.
• They are chosen for clarity (good semantic contrast and less noise), while other maps can still be informative depending on the image/class.

2. About enlarging a 1/16 super-layer heat map and projecting it onto the full image:

• The standard approach is:
  • aggregate the feature map (e.g., channel mean),
  • normalize to 0,1,
  • resize to original image size (we use bicubic interpolation),
  • apply colormap,
  • blend with the original image (alpha overlay).
• In our implementation we do exactly this with cv2.resize(..., INTER_CUBIC) + weighted overlay, and we export one projection per layer.
  We implemented a full reproducible pipeline here:
• Feature-map projection script (all layers):
  https://github.com/ahmedelbamby-aast/yolov13-custom/blob/main/kaggle/scripts/37_feature_map_projection.py
• End-to-end run launcher (training + projection artifacts):
  https://github.com/ahmedelbamby-aast/yolov13-custom/blob/main/kaggle/scripts/run_custom_time2_tmp.sh
• Full usage docs (setup/train/val/test/benchmark/custom dataset):
  https://github.com/ahmedelbamby-aast/yolov13-custom/blob/main/kaggle/QUICKSTART.md
  Our run outputs:
• per-layer projected images
• metadata JSON
• a markdown report (ff_maps.md) containing visualizations for every captured layer, so selection is transparent (not limited to 3 examples)