CODEWITHSUNDEEP
algorithmopencvcomputer-vision
Mickey Shine
Mickey Shine

Reputation: 12549

How to improve HOG objector detector for images under different illuminations?

I have some pedestrian images under different illuminations and some of them are very dark and hard to detect. So I am trying to find a way to pre-process these images to improve detection rate. So far I am thinking of:

1) color correction algorithm, such as http://www.ipol.im/pub/art/2011/llmps-scb/ or http://www.ipol.im/pub/art/2011/gl_lcc/

2) OpenCV's CLAHE

But HOG features are 'gradients', kind of 'relative differences'. So are these methods really helpful to improve detection rate when using HOG as features?

Any advices are appreciated!

Upvotes: 2

Views: 884

Answers (1)

killogre
killogre

Reputation: 1800

From my experience with object detection, the kind of problem you describe becomes less of concern when you use the right normalization of the HoG descriptor. I don't know which implementation you are using, so I cannot tell which normalization is used.

If you compute HoG cells for your detection window and normalize the entire descriptor globally (e.g. L2 norm), then you will suffer from illumination variation as you describe, because the areas of the object with large contrast will dominate the areas with low contrast. If instead you normalize each block of cells as suggested by Dalal-Triggs and also by Felzenszwalb et al, then you get local contrast normalization, meaning that a pedestrians half in shadow and half in light will still 'look' like a pedestrian in HoG eyes, every block dominated by the stronger gradients (the silhouette).

If this does not help, you may want to use more modern rigid body detectors, which utilize additional feature channels (e.g. color, texture). They combine HoG with these channels, and use the cascade paradigm to pick the features best suited to your dataset. A good example of this approach is the work of Benenson et al.

Upvotes: 3

Related Questions