About construct_gt_score_maps

[lzw19951010]

Hi, bilylee, thank you for your great work!

I've noticed a slight difference between your implementation and Luca's. I'm not sure whether it is reasonable, it would be great if you can enlighten me on this.

Luca's implementation:

            if dist_from_origin <= rPos
                logloss_label(i,j) = +1;
            else
                if dist_from_origin <= rNeg
                    logloss_label(i,j) = 0;
                else
                    logloss_label(i,j) = -1;
                end
            end

Yours:

y = tf.cast(tf.range(0, ho), dtype=tf.float32) - get_center(ho)
x = tf.cast(tf.range(0, wo), dtype=tf.float32) - get_center(wo)
[Y, X] = tf.meshgrid(y, x)
dist_to_center = tf.abs(X) + tf.abs(Y)  # Block metric
Z = tf.where(dist_to_center <= rPos,
             tf.ones_like(X),
             tf.where(dist_to_center < rNeg,                   
                      0.5 * tf.ones_like(X),
                      tf.zeros_like(X)))

Both of which used balanced weights, but you only generated 1/0.5/0 with your default settings, while Luca's implementation generated 1/0/-1 with respect to different radius. I'm not sure whether this is different by design or simply I'm wrong about it.

[bilylee]

Hi,

I use tf.nn.sigmoid_cross_entropy_with_logits for computing the loss while the original MatConvNet version use something like tf.nn.softmax_cross_entropy_with_logits to compute the loss.

For tf.nn.sigmoid_cross_entropy_with_logits, the label is the probability of the position being the target, therefore, 0 or 1 or 0.5 for not sure. As for tf.nn.softmax_cross_entropy_with_logits, the label is the category of the position belonging to, therefore +1 for target, -1 for background and 0 for not sure.

Please convince yourself these two implementations are identical.

#16 also disccussed this issue in Chinese.

[lzw19951010]

Cool, I did not take the loss function into account. It seems to me that these two implementation are optimizing the score response towards similar targets. However, I still cannot convince myself they are exactly identical. Because in softmax_cross_entropy, score inter-class are non-independent while they are independent in sigmoid. Probably a solid conclusion can be only derived from their backward propagation formulas.

Thank you for your code and your patience!