yolov7添加FPPI评价指标

学术上目标检测大多用mAP去评价一个模型的好坏，mAP用来作为比较模型的指标是挺好的，不过有个问题就是不够直观，比如mAP=0.9到底代表什么呢？平均一个图会误检几个呢？该取什么阈值呢？mAP说明不了，所以有时候我们还需要其他更直观的指标。

FPPI

fppi：false positive per image, 顾名思义就是平均每张图误检的个数。是目标检测中也比较常见的指标。FPPI与missrate(漏检率)可以构成如下的图像，曲线越低越好。

通过FPPI曲线图，我们可以知道在一个FPPI下面的漏检率，可以作为阈值选取的指导。

yolov7中增加FPPI

FPPI实现

yolo7中的评价指标实现位于utils/metrics.py中，我们只需要参照mAP指标在其中增加FPPI的内容即可：

def fppi_per_class(tp, conf, pred_cls, target_cls, image_num, plot=False, save_dir='.', names=(), return_plt=False):
    """ Compute the false positives per image (FPPW) metric, given the recall and precision curves.
    Source:
    # Arguments
        tp:  True positives (nparray, nx1 or nx10).
        conf:  Objectness value from 0-1 (nparray).
        pred_cls:  Predicted object classes (nparray).
        target_cls:  True object classes (nparray).
        plot:  Plot precision-recall curve at [email protected]
        save_dir:  Plot save directory
    # Returns
        The fppi curve 
    """
    # Sort by objectness
    i = np.argsort(-conf)
    tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]

    # Find unique classes
    unique_classes = np.unique(target_cls)
    nc = unique_classes.shape[0]  # number of classes, number of detections

    # Create Precision-Recall curve and compute AP for each class
    px, py = np.linspace(0, 1, 1000), np.linspace(0,100,1000) # for plotting
    r = np.zeros((nc, 1000))
    miss_rate = np.zeros((nc, 1000))
    fppi = np.zeros((nc, 1000))
    miss_rate_at_fppi = np.zeros((nc, 3)) # missrate at fppi 1, 0.1, 0.01
    p_miss_rate = np.array([1, 0.1, 0.01])
    for ci, c in enumerate(unique_classes):
        i = pred_cls == c
        n_l = (target_cls == c).sum()  # number of labels
        n_p = i.sum()  # number of predictions

        if n_p == 0 or n_l == 0:
            continue
        else:
            # Accumulate FPs and TPs
            fpc = (1 - tp[i]).cumsum(0)
            tpc = tp[i].cumsum(0)

            # Recall
            recall = tpc / (n_l + 1e-16)  # recall curve
            r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0)  # negative x, xp because xp decreases
            miss_rate[ci] = 1 - r[ci]

            fp_per_image = fpc/image_num
            fppi[ci] = np.interp(-px,-conf[i], fp_per_image[:,0], left=0)

            miss_rate_at_fppi[ci] = np.interp(-p_miss_rate, -fppi[ci], miss_rate[ci])
    
    if plot:
        fig = plot_fppi_curve(fppi, miss_rate, miss_rate_at_fppi, Path(save_dir) / 'fppi_curve.png', names)

    if return_plt:
        return fppi, miss_rate, miss_rate_at_fppi, fig

    return miss_rate, fppi, miss_rate_at_fppi

和mAP比较类似
$fppi=fp/image\_num$
$missrate=1-recall$
将fppi以对数坐标画图：

def plot_fppi_curve(px,py, missrate_at_fppi, save_dir='fppi_curve.png', names=()):
    fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
    py = np.stack(py, axis=1)
    # semi log
    for i, y in enumerate(py.T):
        ax.semilogx(px[i],y, linewidth=1, label=f'{names[i]} {missrate_at_fppi[i].mean():.3f}')  # plot(recall, precision)
    
    ax.semilogx(px.mean(0), py.mean(1), linewidth=3, color='blue', label='all classes %.3f' % missrate_at_fppi.mean())

    ax.set_xlabel('False Positives Per Image')
    ax.set_ylabel('Miss Rate')
    ax.set_xlim(0, 100)
    ax.set_ylim(0, 1)
    ax.grid(True)
    plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
    fig.savefig(Path(save_dir), dpi=250)

    return fig

训练中调用

在test.py中在map计算的下方增加fppi的计算：

p, r, f1, mp, mr, map50, map, t0, t1, mfppi_1 = 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
........
........
stats = [np.concatenate(x, 0) for x in zip(*stats)]  # to numpy
if len(stats) and stats[0].any():
	p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names)
	miss_rate, fppi, miss_rate_at_fppi = fppi_per_class(*stats, plot=plots, image_num= image_num, save_dir=save_dir, names=names)
	mfppi_1 = miss_rate_at_fppi[:,0].mean()
	ap50, ap = ap[:, 0], ap.mean(1)  # [email protected], [email protected]:0.95
	mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean()
	nt = np.bincount(stats[3].astype(np.int64), minlength=nc)  # number of targets per class
else:
	nt = torch.zeros(1)# Print results
pf = "%20s" + "%12i" * 2 + "%12.3g" * 5  # print format
print(pf % ("all", seen, nt.sum(), mp, mr, map50, map, mfppi_1))
········
#返回fppi
return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist(), mfppi_1), maps, t

wandb中增加，我习惯使用wandb, 所以加了这步，如果不使用wandb的话，上面的函数不返回fppi就不用修改train.py了：

train.py# Log
            tags = [
                "train/box_loss",
                "train/obj_loss",
                "train/cls_loss",  # train loss
                "metrics/precision",
                "metrics/recall",
                "metrics/mAP_0.5",
                "metrics/mAP_0.5:0.95",
                "val/box_loss",
                "val/obj_loss",
                "val/cls_loss",  # val loss,
                "val/missrate@fppi=1",
                "x/lr0",
                "x/lr1",
                "x/lr2",
            ]  # params

效果

训练过程中会在mAP的后面打印fppi, 训练完成后以及调用test.py测试时，会画fppi图：

结语

本文简述了在yolov7中增加FPPI评价指标，可以用来直观的表现模型的效果，指导阈值的选取。