Drag the box to move it, or grab an edge or corner to resize it. The proposal is measured in whole image pixels; the feature map carries one cell per 16.96 pixels, so the projected box lands between cells. The map is 76 cells wide but only 74.5 of those columns sit over the photograph: the network resizes the image and then pads it out to a multiple of 32 before it convolves, and the grey band on the right of the map is that padding. RoI pooling rounds the projected box to whole cells — the shaded sliver is what that rounding adds or drops — and then rounds each of the 49 bin boundaries to whole cells as well, which is why neighbouring bins end up reading some of the same cells. Click a bin in the pooled grid to see which cells it actually reads. Switch to RoI align and nothing is rounded: the bins sit where the arithmetic puts them, and each is sampled at four points interpolated from the four nearest cell centres. Pooling takes the maximum over whole cells; align averages its four samples, which is what torchvision's roi_align does, so the two grids differ in value as well as in placement. Edges round to the nearest cell boundary, as the paper describes; implementations differ over whether the last cell is included.