For calibration, it’s not advised to use many images. From past experiece 64 images is a good size of dataset for calibration. The reason is, that during the calibration profcess the qunatizer collects max/min/zero-point data of each layer. Using large calibration set pushes the max and the min farther away, thus makeing more outlier nodes. This in turn force a large dynamic range that increases the quantization noise.