pico
pico copied to clipboard
nenadmarkus
model can't converge
Hi there, I'm trying to use picolrn to train my own face detectors (using AFLW dataset). first i use genki.py to preprocess my positive samples(24384faces),then i choose around 15G background images. When i train the model, I haven't been able to get it to converge.In the first stage,there are 4 trees.In every stage,there are max trees as picolrn has setted. With this in mind, I was wondering whether you had any advice as to how to solve this problem? Sorry for my poor English! Thanks!
Is the default training sample (GENKI) working well? I.e., were you able to get a working cascade?
If so, the dataset you use may be too "diverse" for trees of depth 5 (default). Try increasing the tree depth (line 851 in picolrn.c). E.g., set tdepth = 7.
Thanks for answering my question The default training sample(GENKI) is working well(i mean about training).As you know,the GENKI dataset has a few people face,so the detection performance is not good in FDDB dataset. I have read your paper.In the paper,your training sample is chosen in AFLW dataset，and setting the depth of trees 6.I also use AFLW dataset and set the depth of trees 6.And i haven't changed any other parameters. Could you tell me more details about how to choose the positive and negative samples(especially in quantity) Buy the way,only setting tdepth=12,in the first stage,the number of tree is 1.
Did you check whether your data-generating script outputs good images (faces) and that the bounding box coordinates are OK? Can you send me some face images with drawn bounding boxes generated by the script?
Thanks for answering my question
I have checked this question.That's no problem(i think).I will send you some face images with drawn bounding boxes generated by the script through the annex(especially i picked some side face).I hope you can give me some advice about the training samples.
At 2017-04-11 19:25:54, "Nenad Markuš" [email protected] wrote:
Did you check whether your data-generating script outputs good images (faces) and that the bounding box coordinates are OK? Can you send me some face images with drawn bounding boxes generated by the script?
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.
I have checked this question.That's no problem(i think).I will send you some face images with drawn bounding boxes generated by the script through the annex(especially i picked some side face).I hope you can give me some advice about the training samples.
The images seem OK. Note that pico will have problems if there's too much variation in the dataset.
Can you send me the output of picolrn.c when run on your data?
myfacefinder.zip OK.I used the entire AFLW dataset(24384 face)，and used genki.py to process the image(just nrands set to 3 ).Then i got 1584 trees.I have describe the problem when i first ask you this question.
myfacefinder.zip
I expected the text that picolrn.c outputs to the console. Please send that.
Could you please give me your training data?
The original facefinder cascade was produced in 2013. Thus, I do not have its training data any more.
I get the same problem. I am willing to train a model, I have the right train data, but I do not know how to use the "picolrn.c". I have a puzzle about the function——"main"——in picolrn.c.
Maybe you need to read the code, argv[1] means the training data, argv[2] means the output dir. You can find these in the founction intlearn_with_default_parameters(char* trdata, char* dst)
At 2017-12-15 14:15:01, "长牙的小蝌蚪" [email protected] wrote:
I get the same problem. I am willing to train a model, I have the right train data, but I do not know how to use the "picolrn.c". I have a puzzle about the code——"main"——in picolrn.c.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.