map50和mpa50-95

### MAP50 和 MPA50-95 的定义

MAP50（Mean Average Precision at IoU threshold of 50%）是指在目标检测任务中，当预测框与真实框之间的交并比（Intersection over Union, IoU）达到至少50%时计算得到的平均精度均值[^1]。

MPA50-95（Mean Average Precision from IoU thresholds ranging from 50% to 95%, step by 5%），则是在多个IoU阈值范围内评估模型性能的一种方式。具体来说，在这个度量标准下会考虑从50%到95%，每隔5个百分点作为一个独立的IoU阈值来分别计算AP（Average Precision），最后取这些不同阈值下的AP的平均值得到最终的结果。

### 计算方法对比

对于MAP50而言，只需要关注单个固定的IoU=0.5的情况即可完成评价过程；而MPA50-95则是综合考量了一系列不同的IoU水平上的表现情况，因此能够更加全面地反映模型在整个范围内的鲁棒性和准确性。

def calculate_map_50(precision_recall_curve):
    """
    Calculate the mean average precision at IOU threshold of 50%
    
    Args:
        precision_recall_curve (list): List containing tuples of (precision, recall)

    Returns:
        float: Mean Average Precision @ .5IOU
    """
    ap = sum([p * r for p, r in precision_recall_curve]) / max(len(precision_recall_curve), 1e-6)
    return ap


def calculate_mpa_50_to_95(all_precision_recall_curves):
    """
    Calculate the mean average precision across multiple IOU thresholds
    
    Args:
        all_precision_recall_curves (dict): Dictionary with keys as iou_thresholds and values as lists of (precision, recall) pairs.

    Returns:
        float: Mean Average Precision @[.5:.05:.95]
    """
    ious = [round(x*0.05,2) for x in range(10,20)]
    aps = []
    for iou in ious:
        pr_curve = all_precision_recall_curves[iou]
        ap = sum([p*r for p,r in pr_curve])/max(len(pr_curve), 1e-6)
        aps.append(ap)
        
    return sum(aps)/len(ious)