Face R-CNN論文解讀

網路結構：

這篇文章在基於faster rcnn做了一些改進：

First, based on a newly developed loss function called center loss [33], we design a new multi-task
loss function in the Fast R-CNN model to supervise the learning of discriminative deep features for
face/non-face classification.

Second, in the learning of deep features, we use online hard example
mining algorithm [1] to generate hard samples (in which the ratio of positive samples to negative
samples is set to be 1:1) for subsequent processing.

Third, we use multi-scale training strategy to
help improve the detection performance.

這篇文章與faster rcnn的不同之處主要在fast rcnn部分

1、Center Loss

fast rcnn的損失函式變為：

為了使得center loss均衡，一個mini batch中正負樣本比例限制為1：1.

2、Online Hard Example Mining

每次從正負樣本中各選出loss最大的N個樣本加入下次訓練。

3、Multi-Scale Training

Instead of using a fixed scale for all the training images in the typical Faster R-CNN framework, we
design a multi-scale representation for each image by resizing the original image to different sizes
during the training process. In this way, the learned model is more adapted to low-resolution faces.
In the testing process, multi-scale testing is performed accordingly, and the predicted bounding
boxes at different image scales are combined into the final output.

為了弱化尺度影響(或者更好地檢測小目標)，訓練階段圖片會經過不同尺度縮放。

We train the detector using VGG19 with the ImageNet pre-trained model.