Mastering Computer Vision with TensorFlow 2.x

SiamMask-based tracking

At CVPR 2019, Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Phillip H.S.Torr came up with SiamMask in their paper Fast Online Object Tracking and Segmentation: A Unifying Approach. More details about the paper can be found at https://arxiv.org/abs/1812.05050.

SiamMask uses a single bounding box initialization and tracks object bounding boxes at 55 frames per second.

Here, the simple cross-correlation of the Siamese network is replaced by a depth-wise correlation to produce a multi-channel response map:

• A w x h binary mask (one for each feature map) is designed using a simple two-layer 1 x 1 convolutional neural network, hf. The first layer has 256 channels and the second layer has 63 x 63 channels.
• ResNet-50 ...