OpenFace/matlab_version/experiments_JANUS

Scripts for evaluating CE-CLM, and CLNF methods on the IJB-FL dataset - http://www.bmva.org/bmvc/2016/papers/paper095/paper095.pdf

To run the models on the data using models trained on 300W:
- Script_CLNF_wild.m

To run the models on the data using models trained on 300W + MultiPIE:
- Script_CECLM_general.m
- Script_CLNF_general.m

To run CE-CLM model trained on 300W + MultiPIE + Menpo:
- Script CECLM_menpo.m

To visualize the results of the models against baselines (using inner face markup - 49, and full face markup - 68):
Display_ceclm_results_49.m
Display_ceclm_results_68.m

To construct the error table:
Construct_error_table.m

The 300W trained CLNF model performs best so it is used in comparisons, for CE-CLM 300W + MultiPIE works better so it is used. We do not use Menpo trained CE-CLM model for comparison as other models have not been trained on the more extensive Menpo data, but it is the best performing model.

For a fair comparison all the models and baselines are initialized using the bounding boxes generated by adding noise to ground truth bounding boxes.