search

Tobias-Fischer / rt_gene

RT-GENE: Real-Time Eye Gaze and Blink Estimation in Natural Environments
http://www.imperial.ac.uk/personal-robotics
Other
365 stars 68 forks source link

reproduce #46

Closed FigaroK closed 4 years ago

FigaroK commented 4 years ago

Hi, I ran your model training code with only changing the path. And I get following results: Ensemble: 13.480204036924533 +- 3.772117425419047 Model 0: 15.14554381498352 +- 3.4026110207019453 Model 1: 14.49058686821486 +- 3.9057199613231455 Model 2: 13.60103630178998 +- 3.4098169707825843 Model 3: 14.496607138990932 +- 3.3971626371824732 Ensemble: 12.86649487961544 +- 3.6739599920738577 Model 0: 13.41469958155073 +- 3.516076889661513 Model 1: 13.237533686031147 +- 3.5218109462077685 Model 2: 13.376636419685994 +- 3.341763248689219 Model 3: 13.861938083633717 +- 3.30115446948625 Ensemble: 12.639686620568382 +- 3.6341994290148607 Model 0: 13.336198507628263 +- 3.2624630095315488 Model 1: 13.110821038670409 +- 3.4787565847257684 Model 2: 13.355155726032125 +- 3.0450093915309906 Model 3: 13.533675092731762 +- 3.055964720850351 Ensemble: 13.048184325211764 +- 3.711550885325419 Model 0: 13.621509926156996 +- 3.2997626846039445 Model 1: 13.280787940634553 +- 3.5138811741676848 Model 2: 13.611282873524203 +- 3.4179222462072456 Model 3: 13.632872017170953 +- 3.4504936315324413

Is this a roughly correct result? Your paper reported the within-dataset result of rt-gene is 7.7 degrees. what did i ignore?

ahmed-alhindawi commented 4 years ago

Hi, There's been significant developments on the repo since the paper and the model hasn't been retrained. We now use a different face detector and a face landmark extractor, the latter specifically affects performance as the eye patches extracted aren't the same.

If you checkout the repository from the first commit, you'll be able to recreate the results of the paper.

We're in the process of retraining the gaze estimation network for the new developments - keep an eye out!

FigaroK commented 4 years ago

Hi, The code I used is exactly at the first commit (git log id: f457c1b). I got the results of 3-fold cross validation as above, and it seems far from 7.7 degrees. Could you provide some help? My tensofrflow is 1.5.0 and keras is 2.1.4 .

ahmed-alhindawi commented 4 years ago

Sorry for the delay in writing back - I've managed to re-create RT-GENE in pytorch; have a look at the pytorch_training branch. Perhaps the code there might help you? The training there regularly goes down to 7 +/- 2 degrees within 3 epochs