please-rewrite-and-retouch-these-paragraphs-nHRE-NDC-adopts-a-multi-scale-structure-which-is-shown-in-the-figure-with-3-layers-as-an-example-Different-from-some-prior-multi-scale-networks-used-in-some-classification-or-recognition-tasks-they-often-ex

The HRE-NDC network has a multi-scale structure, as depicted by the example figure with three layers. Unlike other multi-scale networks used for classification or recognition tasks, which extract features at multiple scales and concatenate them, HRE-NDC aims to minimize computation at high resolution scales for efficiency. To achieve this, the ground truth is first downsampled to various scales to generate corresponding pseudo targets. Each downsampling reduces the grid resolution by half, and the lowest scale in the figure has been downsampled twice (j=2). The generated pseudo targets can be represented as follows: \n