Dip in the end part of PR curve

[PawelLawrynowicz]

I'm aware that the coco_eval.eval['precision'] has the [TxRxKxAxM] format. Therefore below code should generate 3 PR curves for 3 IoU thresholds as mentioned here. The results are bizzare. After certain recall threshold both metrics drop to 0.

all_precision = coco_eval.eval['precision']
    pr_50 = all_precision[0, :, 0, 0, 2]  # [email protected]
    pr_75 = all_precision[5, :, 0, 0, 2]  # [email protected]
    pr_95 = all_precision[9, :, 0, 0, 2]  # [email protected]

    x = np.arange(0, 1.01, 0.01)
    ax = plt.subplot(111)

    ax.plot(x, pr_50, label='[email protected]')
    ax.plot(x, pr_75, label='[email protected]')
    ax.plot(x, pr_95, label='[email protected]')
    ax.xaxis.set_major_locator(plt.MultipleLocator(0.25))
    ax.yaxis.set_major_locator(plt.MultipleLocator(0.25))
    ax.grid()
    ax.spines[['right', 'top']].set_visible(False)

    plt.xlabel('Recall')
    plt.ylabel('Precision')

    plt.legend()

Here is the result plot for yolox_nano for class 'person':

[iamsiddhantsahu]

@PawelLawrynowicz I also this same question