[target tracking] |siamfc

The paper ：Fully-Convolutional Siamese Networks for Object Tracking
Code ：https://github.com/huanglianghua/siamfc-pytorch

We trained a twin network to locate sample images in larger search images . A further contribution is the novel twin architecture of complete convolution relative to the search image ： The dense and effective sliding window evaluation is realized by calculating the bilinear layer of the cross-correlation between its two inputs .

Z Templates ,x Search area ,ϕ Transform function Feature mapping operation ,g Is a simple distance or similarity measure


* Represents convolution operation , Get one 17^17 Of score map, Represents the search region The similarity value between each position in the template .
take 17×17 The score graph of is generated by bicubic interpolation 272×272 Image , To determine the position of an object .

The algorithm itself is to compare the similarity between the search area and the target template , Finally get the search area score map. In fact, in principle , This method is very similar to the method of correlation filtering . It matches the target template one by one in the search area , This method of calculating similarity by translation matching one by one is regarded as a convolution , Then find the point with the largest similarity value in the convolution result , As the center of new goals .

The picture above ϕ It's actually CNN Part of , And two ϕ The network structure is the same , This is a typical twin neural network , And in the whole model only conv Layer and the pooling layer , So this is also a typical full convolution （fully-convolutional） neural network .

Extract the sample image and search image of each frame offline , To avoid resizing the image during training .

Feature extraction network φ

φ The corresponding feature extraction network adopts AlexNet

Input

about z
The image cutting method is to select a tight bounding box ( w x h ) , And select more pixels in a certain range around p As the context boundary , p The calculation method of is p = ( w + h ) / 4 , Multiply by a scale parameter , become A ( 127 x 127 )

about x
Will be cut out from the whole picture 255×255 Pictures of the , The center of clipping is predicted by the previous frame bounding-box Center of .

When template image z   And search for images x When not enough cutting , The missing part will be averaged RGB Value padding .

Loss function