If by resize you mean, reduce dimensionality of your data, then you can use PCA or similar approaches based on Eigen values and eigen vectors to use your data efficiently with less complexity. and you can use only useful features without loosing lots of representation information. I am adding some links of papers you may review to get an idea.
PCA is good for dimensionality reduction, but SVD is also used for same purpose that give the diagonal matrix. you can reduce the dimensionality by selecting data according to eigen values. (Higher eigen value data will contain more useful information then lower eigen value data).
I guess that you mean to scale the images and not the data. In the INRIA dataset you have the frames with the annotations but also you have the cropped pedestrias at 64x128 + an extra margin. If you want to scale them to a different scale you should crop them from the original image and scale it using bilinear interpolation which gives the best results vs speed.