SSD prior box creation

create_prior_box.py

import pickle
import numpy as np
import pdb

img_width, img_height = 300, 300
box_configs = [
    {'layer_width': 38, 'layer_height': 38, 'num_prior': 3, 'min_size':  30.0,
     'max_size': None, 'aspect_ratios': [1.0, 2.0, 1/2.0]},
    {'layer_width': 19, 'layer_height': 19, 'num_prior': 6, 'min_size':  60.0,
     'max_size': 114.0, 'aspect_ratios': [1.0, 1.0, 2.0, 1/2.0, 3.0, 1/3.0]},
    {'layer_width': 10, 'layer_height': 10, 'num_prior': 6, 'min_size': 114.0,
     'max_size': 168.0, 'aspect_ratios': [1.0, 1.0, 2.0, 1/2.0, 3.0, 1/3.0]},
    {'layer_width':  5, 'layer_height':  5, 'num_prior': 6, 'min_size': 168.0,
     'max_size': 222.0, 'aspect_ratios': [1.0, 1.0, 2.0, 1/2.0, 3.0, 1/3.0]},
    {'layer_width':  3, 'layer_height':  3, 'num_prior': 6, 'min_size': 222.0,
     'max_size': 276.0, 'aspect_ratios': [1.0, 1.0, 2.0, 1/2.0, 3.0, 1/3.0]},
    {'layer_width':  1, 'layer_height':  1, 'num_prior': 6, 'min_size': 276.0,
     'max_size': 330.0, 'aspect_ratios': [1.0, 1.0, 2.0, 1/2.0, 3.0, 1/3.0]},
]
variance = [0.1, 0.1, 0.2, 0.2]
boxes_paras = []

def create_prior_box():
    for layer_config in box_configs:
        layer_width, layer_height = layer_config["layer_width"], layer_config["layer_height"]
        num_priors = layer_config["num_prior"]
        aspect_ratios = layer_config["aspect_ratios"]
        min_size = layer_config["min_size"]
        max_size = layer_config["max_size"]

        step_x = float(img_width) / float(layer_width)
        step_y = float(img_height) / float(layer_height)

        linx = np.linspace(0.5 * step_x, img_width - 0.5 * step_x, layer_width)
        liny = np.linspace(0.5 * step_y, img_height - 0.5 * step_y, layer_height)

        centers_x, centers_y = np.meshgrid(linx, liny)
        centers_x = centers_x.reshape(-1, 1)
        centers_y = centers_y.reshape(-1, 1)

        assert(num_priors == len(aspect_ratios))
        prior_boxes = np.concatenate((centers_x, centers_y), axis=1)
        prior_boxes = np.tile(prior_boxes, (1, 2 * num_priors))

        box_widths = []
        box_heights = []
        for ar in aspect_ratios:
            if ar == 1 and len(box_widths) == 0:
                box_widths.append(min_size)
                box_heights.append(min_size)
            elif ar == 1 and len(box_widths) > 0:
                box_widths.append(np.sqrt(min_size * max_size))
                box_heights.append(np.sqrt(min_size * max_size))
            elif ar != 1:
                box_widths.append(min_size * np.sqrt(ar))
                box_heights.append(min_size / np.sqrt(ar))
        box_widths = 0.5 * np.array(box_widths)
        box_heights = 0.5 * np.array(box_heights)

        # Normalize to 0-1
        prior_boxes[:, ::4] -= box_widths
        prior_boxes[:, 1::4] -= box_heights
        prior_boxes[:, 2::4] += box_widths
        prior_boxes[:, 3::4] += box_heights
        prior_boxes[:, ::2] /= img_width
        prior_boxes[:, 1::2] /= img_height
        prior_boxes = prior_boxes.reshape(-1, 4)
        # clip to 0-1
        prior_boxes = np.minimum(np.maximum(prior_boxes, 0.0), 1.0)
        piror_variances = np.tile(variance, (len(prior_boxes),1))
        boxes_para = np.concatenate((prior_boxes, piror_variances), axis=1)
        boxes_paras.append(boxes_para)

    return np.concatenate(boxes_paras, axis=0)


if __name__ == "__main__":
    boxes_paras = create_prior_box()
    priors = pickle.load(open('../Datasets/ExtraData/prior_boxes_ssd300.pkl', 'rb'))
    diff = boxes_paras - priors
    # pdb.set_trace()
    print("simi {}, max value {}, min value {}".format(diff.shape, diff.max(), diff.min()))

What kind of box_configs are used for the 500x500 version of the model in the SSD paper?
If you change image_width and img_height, the size of the array is still same (7308,8)

Curious about the 512x512 too? found SSD from balancap which specifies some parameters but not all

Many thanks it is really helpfull.
May you clarify how you selected min_size and max_size for every feature maps ? In the original paper I see s_min=0,2 and s_max=0,9 but I do not understand how those are linked to your setting.

This is really helpful
But have a doubt the num_prior in box_configs refers to the number of prior box in each feature layer right ?
If so [4, 6, 6, 6, 4, 4] is this the correct number of boxes in each layer rather than [3, 6, 6, 6, 6, 6]
can you tell me why you used [3, 6, 6, 6, 6, 6] this ?

In box_configs variable, how the min and max sizes are defined?