preview

Analysis And Analysis Of The Fast R-CNN Model

Decent Essays
Open Document

Fig. \ref{fig:FastRCNN} shows the Fast R-CNN model. The input to fast R-CNN is an entire image along with object proposals. In the first stage, convolutional feature maps are extracted when the image is passed through different convolutional and max pooling layers (Module A in Fig. \ref{fig:FastRCNN}). For each object proposal, a feature vector of fixed size is then extracted from the feature map by a Region of Interest (RoI) pooling layer (Module B). The role of the RoI pooling layer is to convert the feature, in a valid RoI, into small feature maps of fixed size ($X \times Y$ e.g., 7 $\times$ 7), using max pooling. $X$ and $Y$ are the layer hyper-parameters. A Region of Interest itself is a rectangular window which is characterized by a …show more content…

Each model consists of 5 max pooling layers, and 5 and 13 convolutional layers. During initialization, a Fast R-CNN model undergoes three transformations. \textbf{First}, a RoI pooling layer (Module B) replaces the last max pooling layer to make it compatible with the network's first fully connected layer (e.g., for VGG16 $X$ = $Y$ = 7). \textbf{Second}, the two sibling layers ((Module C and D), discussed above, replace the last fully connected layer and softmax of the network. \textbf{Third}, the network model is tweaked to accept two data inputs i.e., a list of images and their RoIs. \subsubsection{Fast R-CNN Training} Stochastic Gradient Descent (SGD) in mini batches (details provided in Sec. 5.3.3) is used to train fast R-CNN by sampling $I$ images and $R/I$ RoIs from each image. As noted by the authors, a small $I$ helps in decreasing the mini-batch computational complexity. For instance, if $I$ = 2 and $R$ = 128, the training strategy has been found to be 64 times faster than sampling one RoI from 128 images. In addition, the training process of the Fast R-CNN is further streamlined with one fine-tuning stage that not only optimizes the softmax classifier but also the bounding-box regressors, and the SVMs \citep{he2014spatial}. \subsubsection{Detection using Fast R-CNN} For detection, the fast R-CNN takes as an input, an image along with a list of $R$ object proposals to score. During testing, $R$ is kept at around 2000.

Get Access
Related
Get Access