21_dec_2016

Finding Tiny Faces

Questions:

@edgarriba: How do they calculate the confidence for each detected face?

The the predicted heatmap at a pixel position (x, y) specifies the confidence of a fixed size detection centered at (x, y).

@gitabcworld: At the general arquitecture schema (Figure 9), they train first a FCN to calculate the initial heatmap with only 40-140 bounding box faces (which corresponds to 1x scale). Afterwards they train 3 differents CNN starting from this shared initial FCN? or do they use only a unique FCN?

They use only a unique FCN. However, during the training it recieve images of different scales.

@prlz77: How to interpret figure 10? Is a template a fine-tuned CNN?

Deadline 21/12/2016 14:30