Closed bushou-yhh closed 2 years ago
我们分别计算了两个分类器在每一个锚框上的预测差异(即论文中公式(3)的 l_dis
)、图像分类分数(即论文中公式(5)的 {y^}^cls
)与重加权后的预测差异(即论文中公式(7)的 {l~}_dis
),并将这些值对应的锚框用热力图颜色填充,值越大越偏向红色,越小越偏向蓝紫色。最后我们将多个锚框叠加起的热力图与原图按一定比例相加,即可得到论文中的图 6。
We respectively calculated the prediction discrepancy (i.e., l_dis
in the Eq. (3) in the paper), the image classification score (i.e. ,{y^}^cls
in the Eq. (5) in the paper) and the re-weighted prediction discrepancy (i.e., {l~}_dis
in Eq. (7) in the paper) of the two classifiers on each anchor, and fill the anchors corresponding to these values with the color of the heatmap. The larger the value, the more red. The smaller the value, the more blue-violet. Finally, we add the heat map of multiple anchors to the original image in a certain proportion to get Fig. 6 in the paper.