In Figure 7 of your paper, you present t-SNE visualizations of feature distributions for different classes. I would like to clarify the details of the visualization.
Specifically, in Figure 7, does each point represent (1) a pixel from the feature map of size (H, W, C), (2) the entire feature map of a point cloud, or (3) an object-level feature cropped from the feature map?
Specifically, in Figure 7, does each point represent (1) a pixel from the feature map of size (H, W, C), (2) the entire feature map of a point cloud, or (3) an object-level feature cropped from the feature map?