create darknet anchor points using k-means. Uses fast numpy array ops: pascal ~= 1.0 s coco ~= 16 s

k_means_anchor_points.py

# -*- coding: utf-8 -*-
import numpy as np
import xml.etree.ElementTree as ET
from pycocotools.coco import COCO


def convert_coco_bbox(size, box):
    dw = 1. / size[0]
    dh = 1. / size[1]
    x = box[0] + box[2] / 2.0
    y = box[1] + box[3] / 2.0
    w = box[2]
    h = box[3]
    x = x * dw
    w = w * dw
    y = y * dh
    h = h * dh
    return x, y, w, h


def convert_bbox(size, box):
    dw = 1. / size[0]
    dh = 1. / size[1]
    x = (box[0] + box[1]) / 2.0
    y = (box[2] + box[3]) / 2.0
    w = box[1] - box[0]
    h = box[3] - box[2]
    x = x * dw
    w = w * dw
    y = y * dh
    h = h * dh
    return x, y, w, h


def area(x):
    if len(x.shape) == 1:
        return x[0] * x[1]
    else:
        return x[:, 0] * x[:, 1]


def kmeans_iou(k, centroids, points, iter_count=0, iteration_cutoff=25, feature_size=13):

    best_clusters = []
    best_avg_iou = 0
    best_avg_iou_iteration = 0

    npoi = points.shape[0]
    area_p = area(points)  # (npoi, 2) -> (npoi,)

    while True:
        cen2 = centroids.repeat(npoi, axis=0).reshape(k, npoi, 2)
        cdiff = points - cen2
        cidx = np.where(cdiff < 0)
        cen2[cidx] = points[cidx[1], cidx[2]]

        wh = cen2.prod(axis=2).T  # (k, npoi, 2) -> (npoi, k)
        dist = 1. - (wh / (area_p[:, np.newaxis] + area(centroids) - wh))  # -> (npoi, k)
        belongs_to_cluster = np.argmin(dist, axis=1)  # (npoi, k) -> (npoi,)
        clusters_niou = np.min(dist, axis=1)  # (npoi, k) -> (npoi,)
        clusters = [points[belongs_to_cluster == i] for i in range(k)]
        avg_iou = np.mean(1. - clusters_niou)
        if avg_iou > best_avg_iou:
            best_avg_iou = avg_iou
            best_clusters = clusters
            best_avg_iou_iteration = iter_count

        print("\nIteration {}".format(iter_count))
        print("Average iou to closest centroid = {}".format(avg_iou))
        print("Sum of all distances (cost) = {}".format(np.sum(clusters_niou)))

        new_centroids = np.array([np.mean(c, axis=0) for c in clusters])
        isect = np.prod(np.min(np.asarray([centroids, new_centroids]), axis=0), axis=1)
        aa1 = np.prod(centroids, axis=1)
        aa2 = np.prod(new_centroids, axis=1)
        shifts = 1 - isect / (aa1 + aa2 - isect)

        # for i, s in enumerate(shifts):
        #     print("{}: Cluster size: {}, Centroid distance shift: {}".format(i, len(clusters[i]), s))

        if sum(shifts) == 0 or iter_count >= best_avg_iou_iteration + iteration_cutoff:
            break

        centroids = new_centroids
        iter_count += 1

    # Get anchor boxes from best clusters
    anchors = np.asarray([np.mean(cluster, axis=0) for cluster in best_clusters])
    anchors = anchors[anchors[:, 0].argsort()]
    print("k-means clustering pascal anchor points (original coordinates) \
    \nFound at iteration {} with best average IoU: {} \
    \n{}".format(best_avg_iou_iteration, best_avg_iou, anchors*feature_size))

    return anchors


def load_pascal_dataset(datasets):
    name = 'pascal'
    data = []

    for year, image_set in datasets:
        img_ids_filename = '%s/%s/VOCdevkit/VOC%s/ImageSets/Main/%s.txt' % (source_dir, name, year, image_set)
        ifs_img_ids = open(img_ids_filename)
        img_ids = ifs_img_ids.read().strip().split()

        for image_id in img_ids:
            anno_filename = '%s/%s/VOCdevkit/VOC%s/Annotations/%s.xml' % (source_dir, name, year, image_id)
            ifs_anno = open(anno_filename)
            tree = ET.parse(ifs_anno)
            root = tree.getroot()
            size = root.find('size')
            w = int(size.find('width').text)
            h = int(size.find('height').text)

            for obj in root.iter('object'):
                xmlbox = obj.find('bndbox')
                b = (float(xmlbox.find('xmin').text),
                     float(xmlbox.find('xmax').text),
                     float(xmlbox.find('ymin').text),
                     float(xmlbox.find('ymax').text))
                bb = convert_bbox((w, h), b)
                data.append(bb[2:])

            ifs_anno.close()
        ifs_img_ids.close()

    return np.array(data)


def load_coco_dataset(datasets):
    name = 'coco'
    data = []

    for dataset in datasets:
        annfile = '%s/%s/annotations/instances_%s.json' % (source_dir, name, dataset)
        coco = COCO(annfile)
        cats = coco.loadCats(coco.getCatIds())
        base_classes = {cat['id']: cat['name'] for cat in cats}
        img_id_set = set()

        for cat_ids in base_classes.iterkeys():
            img_ids = coco.getImgIds(catIds=cat_ids)
            img_id_set = img_id_set.union(set(img_ids))
        image_ids = list(img_id_set)

        for image_id in image_ids:
            annIds = coco.getAnnIds(imgIds=image_id)
            anns = coco.loadAnns(annIds)
            img = coco.loadImgs(image_id)[0]
            w = img['width']
            h = img['height']

            for ann in anns:
                b = ann['bbox']
                bb = convert_coco_bbox((w, h), b)
                data.append(bb[2:])

    return np.array(data)

if __name__ == "__main__":

    # examples
    # k, pascal, coco
    # 1, 0.30933335617, 0.252004954777
    # 2, 0.45787906725, 0.365835079771
    # 3, 0.53198291772, 0.453180358467
    # 4, 0.57562962803, 0.500282182136
    # 5, 0.58694643198, 0.522010174068
    # 6, 0.61789602056, 0.549904351137
    # 7, 0.63443906479, 0.569485509501
    # 8, 0.65114747974, 0.585718648162
    # 9, 0.66393113546, 0.601564171461

    # k-means picking the first k points as centroids
    img_size = 416
    k = 5

    # change this line to match your system.
    source_dir = "/media/RED6/DATA"

    random_data = np.random.random((1000, 2))
    centroids = np.random.random((k, 2))
    random_anchors = kmeans_iou(k, centroids, random_data)

    subsets = (('2007', 'train'), ('2007', 'val'), ('2012', 'train'), ('2012', 'val'))
    pascal_data = load_pascal_dataset(subsets)
    centroids = pascal_data[np.random.choice(np.arange(len(pascal_data)), k, replace=False)]
    # centroids = pascal_data[:k]
    pascal_anchors = kmeans_iou(k, centroids, pascal_data, feature_size=img_size / 32)

    subsets = ('train2014', 'val2014')
    # subsets = ('test2014', 'test2015')
    coco_data = load_coco_dataset(subsets)
    centroids = coco_data[np.random.choice(np.arange(len(coco_data)), k, replace=False)]
    # centroids = coco_data[:k]
    coco_anchors = kmeans_iou(k, centroids, coco_data, feature_size=img_size / 32)

    print 'done'

Hey hi! I am using your script to generate anchors for PASCAL datasets (2007+2010+train+val) and unfortunately I am unable to reproduce the official anchor values (not even in the ballpark).

I get the following results when I run your script

Average iou to closest centroid = 0.562124231758
Sum of all distances (cost) = 437.875768242
k-means clustering pascal anchor points (original coordinates)     
Found at iteration 2 with best average IoU: 0.568137265025     
[[  0.90712206   5.65400445]
 [  3.02700947   8.69922277]
 [  6.4478616    8.82623828]
 [  7.76996622   2.02551596]
 [ 10.74920055   8.57291079]]

Did I miss something? Any help would be greatly appreciated ! thanks!!

@Kapilkoundinya did you figure it out?

The problem with k-means clustering and IoU is that if you run the algorithm till convergence, it will always put out something like 0.587. So I think to get the promised 0.61 as described in the paper, it is necessary to restart the algorithm as soon as the values start to drop. I ran k-means for 10 min and got as result something like 0.602. So just increase the number of iterations here.

However, there are also other problems like how the authors of the paper implemented their IoU. For example R-CNN https://github.com/rbgirshick/rcnn/blob/master/utils/boxoverlap.m, adds +1 to calculate the width/height. This might not make a big difference, but if there are lots of small things, maybe they will add up and explain why nobody gets 0.61.

@lars76

Should we use absolutes height and width (from Pascal XML) or relative ones (Darkent format)?