A 2016 tracking paper, now in TensorFlow, with a catch

What it does

Implements the SiamFC object tracker from the ECCV 2016 paper “Fully-Convolutional Siamese nets for object tracking,” specifically the improved baseline version from the later CFNet work. You feed it a pretrained network and a video sequence; it tracks an object through frames using correlation-filter-style matching between an exemplar and search regions.

The interesting bit

The authors are upfront about the limitations: this is inference-only, and results will drift “slightly better or worse” from their original MatConvNet implementation. The honesty is refreshing in a field where people usually pretend ports are identical. The setup also still requires Python 2.7, which gives the whole thing a certain archaeological charm.

Key highlights

• Pretrained model available (baseline-conv5_e55.mat from the CFNet release)
• Evaluation script included with VOT2016 sequence support
• Visualization toggle in parameters/run.json
• Direct lineage to well-cited Oxford tracking work (Bertinetto, Valmadre, Torr)
• Academic/educational license; commercial use requires contact

Caveats

• No training: README explicitly states “only allows to use a pretrained net in forward mode”
• Python 2.7 only: virtualenv instructions hardcode /usr/bin/python2.7
• Inference drift expected: authors warn against direct benchmark comparison without using their precomputed MatConvNet results

Verdict

Worth a look if you need a clean, minimal TensorFlow implementation of classic siamese tracking for baselines or teaching. Skip it if you need end-to-end training, modern Python, or production-ready inference.