cfnet

Why we included this project

Visual object tracking in video is genuinely hard, and this is one of the cleaner reference implementations of the correlation-filter approach that carried the field for years. Training the filter end-to-end with a two-layer CNN pulls strong single-object tracking out of a network that runs around 600 KB, which makes it a handy benchmark for anyone on resource-strapped or embedded projects. It's research code through and through: Matlab and MatConvNet, scripts to dig through rather than a plug-and-play package, and the real payoff is walking through the architecture and rerunning the CVPR'17 experiments. If you're building a real-time tracker or comparing lightweight Siamese-style trackers, this is a good baseline to cite and adapt.