Implementation of Bi-level Data Resampling

[SukwonYun]

First of all, great paper! Especially, I enjoyed the way you handled long tail problem via bi-level data resampling.

However, the equation and implemented code came somewhat unclear to me. According to the paper, in Instance-level under-sampling, you had set drop-out rate as follow.

d_i_c = max((r_i - r_c) / r_i * gamma_d, 1.0)

But in the code, it has been written as,
https://github.com/SHTUPLUS/PySGG/blob/59f16b4c49fe1ac424522f5d7f2234cc221c95e9/pysgg/data/datasets/bi_lvl_rsmp.py#L151

https://github.com/SHTUPLUS/PySGG/blob/59f16b4c49fe1ac424522f5d7f2234cc221c95e9/pysgg/data/datasets/bi_lvl_rsmp.py#L154

To sum up, my questions would be,

1. According to the drop out equation written in paper, why we have to use max function? In such case, the drop out rate might be larger than 1 which causes the category in specific image will be all droped out. So, I wonder if the drop-out rate has to be set in range between 0 and 1.

2. Also, according to above drop out equation, _(r_i - r_c / r_i ) * gammad would be always less than 1, so the max function always outputs 1. Thus, also considering question 1, shouldn't the drop-out equation be modified?

3. Lastly, according to the implemented code, what is the role of clip function? Why is it necessary to appear with drop_rate array?

[Scarecrow0]

Hi, Thanks for paying attention to our work.

• First, there is a typo on the CVPR conference version, which the dropout rate is given by the min from 1 and dropout rate: d_{i_c} = min((r_i - r_c) / r_i * gamma_d, 1.0), you can refer on our arxiv version, sorry for confusing you.
• The clip function is used for sampling the relationship by a Bernoulli distribution whose p=dropout rate. The ignored_rel is sampled from the 0-1 uniform distribution. The clip function with the dropout rate as the threshold simulates this sampling process.
• Since we define it as Bernoulli distribution, the p, AKA dropout should be no greater than 1, so there is why we need the min function to handle the boundary condition.