search

machanic / AU_R-CNN

The official implementation code of paper: "AU R-CNN：Encoding Expert Prior Knowledge into R-CNN for Action Unit Detection".
https://arxiv.org/abs/1812.05788
87 stars 24 forks source link

Data Partition List #3

Closed carltaddybearzz closed 4 years ago

carltaddybearzz commented 4 years ago

First of all, thanks for sharing your code!! I have a question about the data list you provided. For BP4D, is the subjects partition same as DRML which detailed here? You are not using the data list provided here, right? Since the later one seems to be unbalanced for different folds. BTW, according to the results from your paper and this one, pre-trained VGG and ResNet can actually achieve ~60 F1 on [BP4D, same partition, frame based detections], which is consistent with my experiments. Originally I thought there were some issues with my implementations since those baselines seem to be as good as some latest conference papers.

machanic commented 4 years ago

@carltaddybearzz 1. I have uploaded my 3-fold data partition into https://github.com/sharpstill/AU_R-CNN/tree/master/necessary_train_files_for_AU_R-CNN/BP4D_3_fold and https://github.com/sharpstill/AU_R-CNN/tree/master/necessary_train_files_for_AU_R-CNN/DISFA_3_fold

  1. Yes, you can check my paper results. Furthermore, I have uploaded all the training necessary files onto https://github.com/sharpstill/AU_R-CNN/tree/master/necessary_train_files_for_AU_R-CNN
machanic commented 4 years ago

Closing now