Merge branch 'master' of https://github.com/MahmudulAlam/Fingertip-Mi…

…xed-Reality

SOLO/README.md

@@ -0,0 +1,65 @@
+## SOLO
+SOLO stands for [```single object localization```](https://github.com/MahmudulAlam/Fingertip-Mixed-Reality/tree/master/SOLO)
+which has been developed for fast and efficient multi-class single object detection.
+In this case, for the purpose of localizing hand for gesture recognition and fingertip detection, it is employed for the purpose of
+hand detection. SOLO divides the input image into grid cells and predicts the probability of having an object in each grid cell. 
+Each grid cell is mapped in the output matrix of the CNN using binary representation. Grid cells that have at least ```50%```
+overlapping with the hand bounding box are labeled as binary :one: and the rest of the cells are labeled as binary :zero:
+
+
+<p align="center">
+ <img src="https://user-images.githubusercontent.com/37298971/65219825-3b5aa500-dadb-11e9-92f3-a1f75eef1b5a.jpg" width="550">
+</p>
+
+## Network Architecture
+For prediction, a fully convolutional network (FCN) is designed based on the common knowledge of the field. The network architecture is
+inspired by the visual geometry group (VGG) model for image classification. 
+
+```
+input = Input(shape=(416, 416, 3))
+
+# Block 01
+x = Conv2D(64, (3, 3), strides=(2, 2), activation='relu', padding='same')(input)
+x = Conv2D(64, (3, 3), activation='relu', padding='same')(x)
+x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(x)
+
+# Block 02
+x = Conv2D(128, (3, 3), activation='relu', padding='same')(x)
+x = Conv2D(128, (3, 3), activation='relu', padding='same')(x)
+x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(x)
+
+# Block 03
+x = Conv2D(256, (3, 3), activation='relu', padding='same')(x)
+x = Conv2D(256, (3, 3), activation='relu', padding='same')(x)
+x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(x)
+
+# Block 04
+x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)
+x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)
+x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)
+x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(x)
+
+# Block 05
+x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)
+x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)
+x = Conv2D(512, (3, 3), activation='relu', padding='same')(x)
+x = Conv2D(1, (1, 1), activation='sigmoid')(x)
+output = Reshape((13, 13), name='output')(x)
+
+model = Model(input, output)
+```
+
+Here, in each of the convolutional layer ```3 x 3``` filters are used and followed by a rectified linear unit (ReLU) activation 
+function except for the final convolutional layer where ```1 x 1``` filter is used to keep the output size same as the desired output
+size and the sigmoid activation function is applied for normalized output.
+
+## Loss Function
+To train the network, following binary cross-entropy loss function is defined where ![H](https://latex.codecogs.com/gif.latex?%7B%5Cmathbb%7BH%7D%7D)
+and ![H^hat](https://latex.codecogs.com/gif.latex?%5Chat%7B%5Cmathbb%7BH%7D%7D) are the ground truth and the predicted
+output matrix of the SOLO. Here, N and M represent the length of the output matrix and the batch size respectively.
+
+<p align="center">
+ <img src="https://user-images.githubusercontent.com/37298971/65953360-9be7cb80-e465-11e9-8e9a-838d9cc75b1c.jpg" width="650">
+</p>
+
+The loss function is optimized using the ADAM optimizer with standard hyperparameter values.