dataset-csv/casme.csv :
- paths for apex frames and its label based on the given mapping (CASME2-coding)
dataset-csv/samm.csv :
- paths for apex frames and its label based on the given mapping (SAMM_Micro_FACS_Codes_v2)
dataset-csv/smic.csv :
- automatically generated by datasetCompiler.ipynb
STEP 1
datasetCompiler.ipynb :
- Create a CSV file of mapping between apex filename and label for SMIC
- Copy apex frames from all datasets into 1 folder (consolidated folder)
- renamed the frames and save the filename and label into cde.csv
STEP 2
datasetCompiler.ipynb :
- Align, crop and resize apex frames to (224,224,3) listed in cde.csv
- The pre-processed frames are copied to the pre-proc folder
STEP 3
datasetCompiler.ipynb :
- Organized the pre-processed frames into subfolders based on its class/labels
- This will make it easier for us to create the training and test sets later
STEP 4
resnet50.ipynb :
- Build resnet50 model with no top and its layers frozen
- Add a Pooling layer and a classifier on top and train this using micro-expression dataset
- Peng et al used resnet-10 with pre-trained weights from ImageNet dataset and then training on macro-expression dataset before fine-tuning on micro-expression datasets -> might try this
STEP 5
efficientnet.ipynb:
- Build EfficientNet-B0, EfficientNet-B1, EfficientNet-B2 model with no top and its layers frozen
- Add a Pooling layer and a classifier on top and train this using micro-expression dataset