OHEM doubt

[vignesh-irtt]

why are you making the loss negative before taking topk losses **vals, = tf.nn.top_k(-neg_conf, k=n_neg)**

[northeastsquare]

neg_conf为背景像素被预测为背景的概率，为1代表确定为预测完全正确；为0代表预测错误，实际为背景像素，却被预测为了不是背景。所以这个值越低，误差越大。 所以，-neg_conf,中的top_k, 为背景像素中，预测误差最大的n_neg个。

[northeastsquare]

实际上，这里可能有个缺陷， pos_pixel_weights_flatten 是gt里面box的像素点生成的，里面小box的pixel weight会大一点，边缘像素也大一点这样 selected_neg_pixel_mask：是背景里面，被预测为背景的像素里面，hardming 出来的。 可是背景里面被预测为正的像素，没有被计算误差。就是下面这句： nets\pixel_link_symbol.py pixel_cls_weights = pos_pixel_weights_flatten + \ tf.cast(selected_neg_pixel_mask, tf.float32)

[vignesh-irtt]

Neg_conf is the probability that the background pixel is predicted as the background, and 1 means that the prediction is completely correct; 0 is the prediction error, which is actually the background pixel, but is predicted to be not the background. So the lower the value, the greater the error. Therefore, top_k in -neg_conf is the n_neg with the largest prediction error in the background pixels.

sorry i thought its as the loss Thank you

[vignesh-irtt]

But as per the paper u haven't selected the negative pixels while calculating the link loss pixel_link_weight *= np.expand_dims(pixel_cls_weight, axis = -1) doesn't seem to contain true for negative pixels pixel_link_label[y, x, n_idx] = 0 does it makes pos cls pixels as neg in link

[northeastsquare]

pixel_link_label[y, x, n_idx] = 0 这句使得，正的像素，周围有点，不是正像素，那么连接为负，