search

symisc / sod

An Embedded Computer Vision & Machine Learning Library (CPU Optimized & IoT Capable)
https://sod.pixlab.io
Other
1.75k stars 213 forks source link

The problem of training samples #5

Closed qaz734913414 closed 6 years ago

qaz734913414 commented 6 years ago

What are the requirements of the SOD training model for samples? How do we deal with the positive and negative samples? Are there any restrictions on the size and so on?

symisc commented 6 years ago

SOD itself with its exported training interfaces does not impose any restrictions on the model you are going to train. The only restriction is how large is your dataset (The larger the dataset is, the more accurate your model) and the amount of RAM, CPU power on the host computer.

qaz734913414 commented 6 years ago

Too many background image collection rounds (> 10K). Perhaps should you stop tra ining and test the accuracy of your model? How to deal with this problem？

symisc commented 6 years ago

This is a hint and no way an error. Two possibility you are facing off:

  1. The negative samples you provided are too small and the cascade is overfitting. It will perform very well on this set but will not generalise on an unknown image set. Provide a larger dataset to fix the issue.
  2. As the message said. Your model perform very well and it's time to stop training and start testing your model in the wild.

-----Original Message----- From: "独自狂战" notifications@github.com To: symisc/sod sod@noreply.github.com Cc: Symisc Systems chm@symisc.net, State change state_change@noreply.github.com Sent: Wed, 13 Jun 2018 3:25 Subject: Re: [symisc/sod] The problem of training samples (#5)

Too many background image collection rounds (> 10K). Perhaps should you stop tra ining and test the accuracy of your model? How to deal with this problem？

-- You are receiving this because you modified the open/close state. Reply to this email directly or view it on GitHub: https://github.com/symisc/sod/issues/5#issuecomment-396793104