search

liuliu / ccv

C-based/Cached/Core Computer Vision Library, A Modern Computer Vision Library
http://libccv.org
Other
7.08k stars 1.72k forks source link

DPMCreate module seems to be stuck in infinite loop #54

Open ashishsharma1512 opened 11 years ago

ashishsharma1512 commented 11 years ago

I am trying to train my own dpm detector using dpmcreate module as mentioned on the http://libccv.org/doc/doc-dpm/

I am using the INRIA dataset, with slightly different positive images(upper body images). The dpmcreate module is running for the last 3 days, but seems to be at the same iteration(as if it is stuck in an infinite loop). All I am seeing for last 3 days is this kind of output(please see attached image):

dpmCreate_stuck_in_infinite_loop

please let me know if i can provide any particular logs/information which might be helpful in understanding the issue better.

liuliu commented 11 years ago

Have you tried to evaluate the intermediate output? (in your output directory, change the starting , to .), it seems like a faulty in the dpmcreate program, but it should produce some recognizable result nevertheless.

ashishsharma1512 commented 11 years ago

thanks for the reply.

sorry, i didn't get you, in my output directory i have couple of intermediate model files (e.g.: model.0.1, model.0.2 etc), do you mean to use them as a final model by changing the extension to .m?

by any chance you will be fixing the fault soon?

liuliu commented 11 years ago

The issue seems to be that even by tuning the threshold to hit 95% tp, there are still fewer negative examples to collect from. I suggest you to check your output/model file on test data to see if it has reasonable output or not, if it has, you can simply just run the post-training step directly (bounding box prediction) and will get your final model.

The model file is not binary, it is a text file, and the beginning "," suggests that it is a work-in-progress one, change it to "." and the dpmdetect will recognize this model. It is always a good practice (at least I do) to check the intermediate model file just to make sure all these parameters you set for training is sane.

Yes, this is a bug in the program, but I won't be able to fix it until mid-Jan (because fixing it requires me to run dpmcreate again to do sanity check, which will conflict with the current face detector training plan), this bug shouldn't interfere much with your training result however.