Weighted Boxes Fusion

paper：https://arxiv.org/pdf/1910.13302.pdf
code: https://github.com/ZFTurbo/Weighted-Boxes-Fusion

前言

我们了解NMS及其soft-NMS。但在这篇博文将简单介绍下WBF(加权框融合)，该方法在开放图像数据集中取得了较大的提高。

WBF的优势

NMS和soft-NMS都排除了一些框，但是WBF利用了所有框的信息。它可以解决某些情况下，模型预测的框不准确。NMS将只能留下一个不准确的框，而WBF将使用来自3个框的信息来修复它，如下图所示，红色的为预测，蓝色的为真值。

代码

# coding: utf-8
__author__ = 'ZFTurbo: https://kaggle.com/zfturbo'


import numpy as np


def bb_intersection_over_union(A, B):
    xA = max(A[0], B[0])
    yA = max(A[1], B[1])
    xB = min(A[2], B[2])
    yB = min(A[3], B[3])

    # compute the area of intersection rectangle
    interArea = max(0, xB - xA) * max(0, yB - yA)

    if interArea == 0:
        return 0.0

    # compute the area of both the prediction and ground-truth rectangles
    boxAArea = (A[2] - A[0]) * (A[3] - A[1])
    boxBArea = (B[2] - B[0]) * (B[3] - B[1])

    iou = interArea / float(boxAArea + boxBArea - interArea)
    return iou


def prefilter_boxes(boxes, scores, labels, weights, thr):
    # Create dict with boxes stored by its label
    new_boxes = dict()
    for t in range(len(boxes)):
        for j in range(len(boxes[t])):
            score = scores[t][j]
            if score < thr:
                continue
            label = int(labels[t][j])
            box_part = boxes[t][j]
            b = [int(label), float(score) * weights[t], float(box_part[0]), float(box_part[1]), float(box_part[2]), float(box_part[3])]
            if label not in new_boxes:
                new_boxes[label] = []
            new_boxes[label].append(b)

    # Sort each list in dict by score and transform it to numpy array
    for k in new_boxes:
        current_boxes = np.array(new_boxes[k])
        new_boxes[k] = current_boxes[current_boxes[:, 1].argsort()[::-1]]

    return new_boxes


def get_weighted_box(boxes, conf_type='avg'):
    """
    Create weighted box for set of boxes
    :param boxes: set of boxes to fuse 
    :param conf_type: type of confidence one of 'avg' or 'max'
    :return: weighted box
    """

    box = np.zeros(6, dtype=np.float32)
    conf = 0
    conf_list = []
    for b in boxes:
        box[2:] += (b[1] * b[2:])
        conf += b[1]
        conf_list.append(b[1])
    box[0] = boxes[0][0]
    if conf_type == 'avg':
        box[1] = conf / len(boxes)
    elif conf_type == 'max':
        box[1] = np.array(conf_list).max()
    box[2:] /= conf
    return box


def find_matching_box(boxes_list, new_box, match_iou):
    best_iou = match_iou
    best_index = -1
    for i in range(len(boxes_list)):
        box = boxes_list[i]
        if box[0] != new_box[0]:
            continue
        iou = bb_intersection_over_union(box[2:], new_box[2:])
        if iou > best_iou:
            best_index = i
            best_iou = iou

    return best_index, best_iou


def weighted_boxes_fusion(boxes_list, scores_list, labels_list, weights=None, iou_thr=0.55, skip_box_thr=0.0, conf_type='avg', allows_overflow=False):
    '''
    :param boxes_list: list of boxes predictions from each model, each box is 4 numbers. 
    It has 3 dimensions (models_number, model_preds, 4)
    Order of boxes: x1, y1, x2, y2. We expect float normalized coordinates [0; 1]
    :param scores_list: list of scores for each model 
    :param labels_list: list of labels for each model
    :param weights: list of weights for each model. Default: None, which means weight == 1 for each model
    :param iou_thr: IoU value for boxes to be a match
    :param skip_box_thr: exclude boxes with score lower than this variable  
    :param conf_type: how to calculate confidence in weighted boxes. 'avg': average value, 'max': maximum value
    :param allows_overflow: false if we want confidence score not exceed 1.0 
    
    :return: boxes: boxes coordinates (Order of boxes: x1, y1, x2, y2). 
    :return: scores: confidence scores
    :return: labels: boxes labels
    '''

    if weights is None:
        weights = np.ones(len(boxes_list))
    if len(weights) != len(boxes_list):
        print('Warning: incorrect number of weights {}. Must be: {}. Set weights equal to 1.'.format(len(weights), len(boxes_list)))
        weights = np.ones(len(boxes_list))
    weights = np.array(weights)

    if conf_type not in ['avg', 'max']:
        print('Unknown conf_type: {}. Must be "avg" or "max"'.format(conf_type))
        exit()

    filtered_boxes = prefilter_boxes(boxes_list, scores_list, labels_list, weights, skip_box_thr)
    if len(filtered_boxes) == 0:
        return np.zeros((0, 4)), np.zeros((0,)), np.zeros((0,))

    overall_boxes = []
    for label in filtered_boxes:
        boxes = filtered_boxes[label]
        new_boxes = []
        weighted_boxes = []

        # Clusterize boxes
        for j in range(0, len(boxes)):
            index, best_iou = find_matching_box(weighted_boxes, boxes[j], iou_thr)
            if index != -1:
                new_boxes[index].append(boxes[j])
                weighted_boxes[index] = get_weighted_box(new_boxes[index], conf_type)
            else:
                new_boxes.append([boxes[j].copy()])
                weighted_boxes.append(boxes[j].copy())

        # Rescale confidence based on number of models and boxes
        for i in range(len(new_boxes)):
            if not allows_overflow:
                weighted_boxes[i][1] = weighted_boxes[i][1] * min(weights.sum(), len(new_boxes[i])) / weights.sum()
            else:
                weighted_boxes[i][1] = weighted_boxes[i][1] * len(new_boxes[i]) / weights.sum()
        overall_boxes.append(np.array(weighted_boxes))

    overall_boxes = np.concatenate(overall_boxes, axis=0)
    overall_boxes = overall_boxes[overall_boxes[:, 1].argsort()[::-1]]
    boxes = overall_boxes[:, 2:]
    scores = overall_boxes[:, 1]
    labels = overall_boxes[:, 0]
    return boxes, scores, labels