Skip to content

thapa-jeevan/Eigen-Face

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Eigen-Face

Setup

Download the following datasets and move into data directory.

  1. training/validation data for PASCAL VOC dataset for non-face categories pretraining
  2. AT&T face images
  3. cropped and aligned celeba faces and face identity
  4. our non-face test images

Also, download our trained checkpoints for the CNN models for face classification and identification and store inside checkpoints directory.

Dimensionality Reduction and Face Reconstruction

Run the following command for analysis on Eigen-Face, importance of principal components and face reconstruction.

python -m src.eigen_face.eigen_face

Face non-face Classification

We build classifiers (LDA, KNN, SVC and RandomForest) to differentiate face images from non-face images by training on a sample of images from AT&T dataset and non-face images created by sampling from PASCAL VOC dataset. PCA is applied to reduce the dimensionality of image representations. To run training and then inference on these datasets, you can run the following command.

python -m src.face_non_face.test_other_classifiers

Also, we train a resnet 18 model for this task. The CNN model is first trained for face non-face classification using face images from CELEBA dataset dataset and non-face images from PASCAL VOC dataset. Then, this model is finetuned in the above mentioned small dataset. To run this pretraining and finetuning pipeline, run the following command.

python -m src.face_non_face.train_resnet

And, run the following to test finetuned model.

python -m src.face_non_face.test_resnet

Face Identification

We approach face identification using multi-class classification and embedding space based models.

First, we train LDA, KNN, SVC and random forest classifiers on AT&T dataset on multi-class classification setting. PCA is applied on the images to reduce the dimensionality. You can run the following command to train and evaluate the pipeline.

python -m src.face_identification.test_other_identifiers

Next, we train resnet 18 using contrastive learning to have faces of same subject to have close embeddings and different subjects to have distant embeddings in cosine space. For this, we pretrain our model on CELEBA dataset and finetune on AT&T dataset. You can run the following command for the overall pretraining and finetuning process.

python -m src.face_identification.train_resnet

And, run the following command to evaluate the finetuned model.

python -m src.face_identification.test_resnet

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages