@xxxfrank Hi, we compute the weights by 1/log(pixel_count).

@PkuRainBow What do the pixel_count mean? Is it the pixel count of the specific class across the dataset? If so, the pixel_count corresponds to loss_weight 0.8373 should be exp(1/0.8373)=3.3013. It's weird!
Correct me if I have anywhere misunderstood.

@eugenelawrence

Thanks for your interest in our work.

We will update the explaination of these weights in our update paper.

Please keep waiting for the arXiv update. We will notice you in the earliest time.

@PkuRainBow
I am wondering would you release the code for calculating the loss weights?

@PkuRainBow
I am wondering would you release the code for calculating the loss weights?

Of course if you need.

@PkuRainBow So, How did you compute the loss weight?

@KeyKy @eugenelawrence @xxxfrank Here we employ the log function firstly,

Then, we will apply L1-normalize over all the log weights w to act as the final weights.

Here I provide the statistics of the pixels' count on the Cityscapes' training set.

Hi, @PkuRainBow
I used the code to Cal. but this is defferent.
`D = [2.01e+9,2.98e+8,9.96e+8,3.39e+7,4.50e+7,6.54e+7,
9.57e+7,2.62e+7,7.21e+8,5.92e+7,1.45e+8,8.21e+7,
1.00e+7,4.13e+8,1.45e+7,1.28e+7,1.45e+7,5.64e+6,
2.57e+7]

w = np.log(D)/np.log(19)
print(w)
w_sum = np.sum(w)
print(w_sum)
w = w/w_sum
print(w)`

@PkuRainBow Could you tell me how to calculate the pixel_count on training dataset?

@PkuRainBow Could you release the code of calculating class_freq in the whole dataset? Thank you.

any news?