If I will test my classifier using images that include the target object in an image among other items. I assume a sliding window is used.
I believe using image segmentation techniques are more effective. Image segmentation techniques such as watershed transformation or region growing are typically used to locate objects and boundaries (lines, curves, etc.) in images. Thresholding technique is another method for segmentation. if a segment below a certain threshold then it is considered a background otherwise it is a foreground segment which must contain an object. Foreground segments are input to an object classifier to decide if this is the target object or not.
To learn about the topic, i would suggest to listen to the lecture (ConvNets for spatial localization, Object detection-lecture 8) in the Stanford deep learning course.
The slides link: http://cs231n.stanford.edu/syllabus.html
The videos link: https://archive.org/details/cs231n-CNNs
For implementation:
This github project provides an example of object detection. They draw a heatmap to indicate the location of the object
https://github.com/heuritech/convnets-keras
In addition, This github page has links of many Deep learning vision apps
https://github.com/kjw0612/awesome-deep-vision
It is good idea that sliding window is used.
With best wishes,
Yoichi Ando
I believe using image segmentation techniques are more effective. Image segmentation techniques such as watershed transformation or region growing are typically used to locate objects and boundaries (lines, curves, etc.) in images. Thresholding technique is another method for segmentation. if a segment below a certain threshold then it is considered a background otherwise it is a foreground segment which must contain an object. Foreground segments are input to an object classifier to decide if this is the target object or not.
- Badges
- Science topic