Ask Your Question
0

How to do hard negative mining for cascade classifier?

asked 2019-11-28 08:05:49 -0600

Zana Zakaryaie Nejad gravatar image

Hi I want to do hard negative mining for my trained cascade classifier. In other words, I want to add false positives to the list of negative images and re-train my cascade to improve accuracy. The question is:

If the cascade detects a large region where a small portion of it is the desired object, then what should I do? The documentation says that negative images must not contain objects. So I shouldn't add it to the negatives. But how to re-train the cascade to ignore such cases?

edit retag flag offensive close merge delete

Comments

Info about negative "magnitude" for real cascades (billions samples):

LBP | vision-ary eyes cascade:

  • 9,000 positive samples
  • 0.7 B of neg. sub-regions
  • Features set: 85.550 features
  • Training time: ~1 days
  • TP: ~ 95.8% of positive training set
  • FN: ~ 04.2% of positive training set
  • FP: ~ 7.51937e-006% of negative training set
  • Training size w=30 h=60 (aspect ratio 1:2)

Full reports:

  1. eyes: http://www.vision-ary.net/2015/11/boo...
  2. faces: http://www.vision-ary.net/2015/03/boo...
  3. cars: http://www.vision-ary.net/2015/06/boo...
  4. pedestrians: http://www.vision-ary.net/2015/03/boo...
ComputerVisionary gravatar imageComputerVisionary ( 2020-03-29 05:37:23 -0600 )edit

1 answer

Sort by ยป oldest newest most voted
1

answered 2019-11-28 10:50:28 -0600

holger gravatar image

updated 2019-11-28 11:04:17 -0600

To clarify - Negatives must not contain the object you want to detect. I can contain other objects (lets call this background). Thats actually good.

So just take the false positives - and use them as negatives. Just make sure the negatives does NOT contain the object you want to detect.

In the case it does - just cut (crop) your object(aka the postive) out.

edit flag offensive delete link more

Question Tools

1 follower

Stats

Asked: 2019-11-28 08:05:49 -0600

Seen: 366 times

Last updated: Nov 28 '19