search

tstandley / image2mass

http://proceedings.mlr.press/v78/standley17a/standley17a.pdf
37 stars 16 forks source link

image2mass 14 geometric properties #11

Open HHLiufighting opened 5 months ago

HHLiufighting commented 5 months ago

Hi, I'm Liuhonghao.

Regarding image2mass, I have two questions for you. First, in the geometry module, you said that 14 features were obtained through mask and length, width and height information, but 35 values were obtained in the code and entered into the model; second The problem is that the density tower has two activation functions. What is the second activation function after the first activation function sigmoid? My understanding is to get a percentage through sigmoid, and then go to the score_to_percentile.pklz sort to find the position of the corresponding percentage. True density value, is my understanding correct?

Thank you for your answer and I look forward to hearing from you.

a4b41f5e9105c871d13842502d3292b

Liuhonghao email: [liuhonghao_q@foxmail.com]

tstandley commented 5 months ago

The 35 values are combinations of the 14. IE a, b, ab, a^2 etc. I really don't remember exactly what I did, but hopefully the details are somewhere in the code.

You're right about the two activations. That's exactly how it works, except I don't think true density is what comes out exactly. There's no supervision for that, we just know that the output of the density tower*the output of the volume tower approximates the mass.