Using 100 classifiers to detect a single class

To solve difficult problems human intuition is not always the best tool.
I will present some of our latest results where we show that:
- There is no need to compute a pixel-wise depth map to find the
distance to pedestrians in stereo images.
- Using 50 classifiers to detect pedestrians is much faster than using
a single classifier (and that the training time does not explode).
- We can train 17 occlusion specific classifiers faster than training
3 classifiers naively (and that 17 classifiers provide better quality
than only 3 or 5).
I will end sketching some of the current research lines and
interesting challenges to tackle.