speculaas
commented
6 years ago Dear Muni,
I just tried a similar classification like what you did.
I have been studying emotion recognition recently.
I also used vggface-2 pretrained model to do feature extraction for 8 emotions as in CK+ dataset.
After classifier training, I ran testing : "python src/classifier.py CLASSIFY" and got "Accuracy: 0.267".
Do you or anybody has any idea why "model = SVC(kernel='linear')" cannot do well in this case?
http://scikit-learn.org/stable/modules/generated/sklearn.svm.SVC.html
Is it because linear classifier not good? Or is it because feature extraction is not good?
I am also trying to fine tune the vgg2-inception-resnet pretrained model using a combination of CK+, JAFFE, TFEID datasets as training set. Then maybe use this fine-tuned inception-resnet to do feature extraction and see what happens.
Also, up till now, I studied the following materials:
https://arxiv.org/abs/1804.08348 This paper gives a survey.
https://arxiv.org/abs/1609.06591 This paper use conv layers of face recognizing net plus a new fully connected layer to do emotion recognition
https://kpzhang93.github.io/papers/icmi.pdf This paper describing their EmotiW 2017 winning method says before fine tuning for group emotion recognition, they first trained Resnet64 on Webface then finetune on ExpW.
BR, Jimmy
Hi, I just ran VGG Face 2 trainer/classifier on the images labeled for 4 emotions, Angry, Happy, Sad and Neutral. Using the split data set I am getting an accuracy of only 25%. Any thought on how I can improve that ?
Also any chance of releasing a inception resnet-v2 frozen model ? I have 2 Pascal GPUs if you need any HW horsepower.
Regards.