Video tracking with deep siamese networks and bayesian optimization

A semiconductor package apparatus comprising:

• one or more substrates;
• and
• logic coupled to the one or more substrates, wherein the logic includes one or more of configurable logic or fixed-functionality hardware logic, the logic coupled to the one or more substrates to:
• select an exemplar crop to be tracked in an initial frame of a video;
• apply Bayesian optimization with each subsequent frame of the video to:
• build a surrogate model of an objective function using Gaussian Process Regression (GPR) based on similarity scores of candidate crops collected from a search space in a current frame of the video;
• and
• determine where to proficiently obtain a next candidate crop in the search space using an acquisition function;
• compare the next candidate crop to the exemplar crop using a Siamese neural network;
• and
• continue comparison of a new candidate crop to the exemplar crop using the Siamese neural network until the exemplar crop has been found in the current frame, wherein the new candidate crop is selected based on an updated surrogate model.

A method of video tracking comprising:

• selecting an exemplar crop to be tracked in an initial frame of a video;
• applying Bayesian optimization with each subsequent frame of the video by:
• building a surrogate model of an objective function using Gaussian Process Regression (GPR) based on similarity scores of candidate crops collected from a search space in a current frame of the video;
• and
• determining where to proficiently obtain a next candidate crop in the search space using an acquisition function;
• comparing the next candidate crop to the exemplar crop using a Siamese neural network;
• and
• continuing comparison of a new candidate crop to the exemplar crop using the Siamese neural network until the exemplar crop has been found in the current frame, wherein the new candidate crop is selected based on an updated surrogate model.