Great idea. However, i tried it on mobilnet-v2 with several images. It does not work well. How can we guarantee the upsampling from feature map to the original image size, the activation map remains spatial correlations (i.e., the area in the original image is correlated with the area of the upsized activation map)? I saw you simply used resize function of CV2.
Great idea. However, i tried it on mobilnet-v2 with several images. It does not work well. How can we guarantee the upsampling from feature map to the original image size, the activation map remains spatial correlations (i.e., the area in the original image is correlated with the area of the upsized activation map)? I saw you simply used resize function of CV2.