no model file

[ardeal]

where is the model file? Do you have the updated final version release?

[luoyetx]

@ardeal Sorry, I don't have a trained model. I may also change the details of the source code. However, the code at this time should work well.

[ardeal]

I should prepare training images and parameters to train a new model, right? According to your experiments, what image datasets are used to train the model? Did you get the same performance with the output of the paper?

[luoyetx]

You should prepare much more data than the paper described. You also need a very powerful machine with many cpu cores to run the algorithm.

Since my lack of these resources, I don't have a well trained model yet. I do collect many training data but don't have a powerful enough machine to handle the data. So, I decide to share the data recently :)

[ardeal]

Could you please share the data format in the following files? how is the data organized in ../data/face.txt and others?

"data": {
    "use_hard": false,
    "face": "../data/face.txt",
    "background": ["../data/hd.txt", "../data/background1.txt", "../data/background2.txt"],
    "test": "../data/test.txt"
},

[luoyetx]

They are described in README

[ardeal]

Hi, in Cart::SplitNodeWithRegression function of cart.cpp file:

int threshold_ = pos_feature_sorted(0, int(pos_n*rng.uniform(0.1, 0.9)));

You got the split threshold by a random way, is this reasonable? or, did you do this according any theory or paper or book?

[ardeal]

I am trying to train the algorithm with 76 pos samples and 1671 neg samples. I merely want to verify whether the code works correctly. but I got the following output all the time: [05/09/16 - 16:44:39] Run out of background images [05/09/16 - 16:44:39] Reset current_idx and restart, reset times = 30 [05/09/16 - 16:44:39] Current augment parameters, should flip = 1, rotation angle = 180

Did you encounter the same issue?

[luoyetx]

1. The threshold in Regression is random. If you iterate all threshold [-255, 255], it will cost too much time, while random threshold can still give you a pretty good result.
2. You need more positive samples, you meet the overfitting problem.