Facial Landmark Detection with Tweaked Convolutional Neural Networks

Published in IEEE Trans. on Pattern Analysis and Machine Intelligence (TPAMI), 2018

Recommended citation: Yue Wu*, Tal Hassner*, KangGeon Kim, Gerard Medioni, and Prem Natarajan. Facial Landmark Detection with Tweaked Convolutional Neural Networks. IEEE Trans. on Pattern Analysis and Machine Intelligence (TPAMI), 40(12):3067--3074, Dec. 2018.

* Denotes joint first authorship / equal contribution

CNN architectures. Left: The vanilla network described in Sec. 2.1 for facial landmark regression. We show that representations extracted from the input to FC5 (marked in red) are highly specialized and reflect facial alignment. Right: Our Tweaked CNN (TCNN) design, diverting intermediate features to K different subsequent, fine-tuned processes in the same dimensions as the original layers.


We present a novel convolutional neural network (CNN) design for facial landmark coordinate regression. We examine the intermediate features of a standard CNN trained for landmark detection and show that features extracted from later, more specialized layers capture rough landmark locations. This provides a natural means of applying differential treatment midway through the network, tweaking processing based on facial alignment. The resulting Tweaked CNN model (TCNN) harnesses the robustness of CNNs for landmark detection, in an appearance-sensitive manner without training multi-part or multi-scale models. Our results on the AFLW, AFW, and 300W benchmarks show improvements over existing work. We further provide results on the Janus benchmark, demonstrating the benefit of our better alignment in face verification.

Old arXiv preprint

Paper on IEEE Xplore



We provide the convolutional neural network models for facial landmark detection. These have been reimplemented in Caffe by Ishay Tubi.

  • A github repository with a Caffe reimplementation of the Vanilla CNN described in the paper. It includes code, prototype files and model weights. Important note: Network weights may still be updated; more accurate networks may be posted here in the future.
  • Vanilla caffe model gist
  • Please see python notebook for example usage.
  • The Vanilla CNN on the Caffe model zoo

We are delighted to announce that our Vanilla CNN for facial landmark detection has been converted to the Wolfram Language Neural Net Framework and is now also available from the Wolfram Neural Net Repository.

This is ongoing work. Please check here regularly for news and updates on the downloads available for this project. If you find our code useful, please add suitable reference to our paper in your work.

What’s New

  • 3-November-2018: Paper officially published on IEEE TPAMI
  • 15-August-2017: Vanilla network for landmark detection now available from the Wolfram Neural Net Repository
  • 3-March-2016: New, slightly more accurate weights have been updated on github for the Vanilla CNN. Vanilla CNN now on Caffe model zooo.
  • 24-February-2016: An initial Vanilla CNN reimplementation is now available.

Copyright and disclaimer
Copyright 2016, Ishay Tubi, Yue Wu and Tal Hassner