Revision history [back]

If it is a consistent error then I would hop through your code and make sure that you do not start drawing rectangles before anything is actually drawn. However, print out the amount of detections returned in negatives, it could well be a false positive getting through, which also occurs in negative images.

Training a pedestrian detector which will get ZERO detections in negatives, is almost impossible unless you have millions of training footage. Structures like light lanterns, trees, sides of houses, have remarkably same features as a human would do.

However adding more knowledge, like background - foreground information, could help you in removing stuff happening in background for example. Give us some more information, some pictures where it goes wrong, your code snippet, ...