Revisiting Facial Keypoint Detection – An Efficient Approach Using Deep Neural Networks

This paper was presented at Big DML, 2021. FKPD_BigDML_Submission-Final.pdf

BigDML_Paper96_FKPD_Final.pptx

Authors - Pratima Dileep, Bharath Kumar Bolla, Sabeesh E

Aim of the Research

The aim of this paper is to design light weight and faster neural networks.
- To evaluate Custom CNN with pre trained architectures
- To evaluate augmentation vs imputation
- To fine tune Custom CNNs manually and by Auto-Keras Tuner

The paper aims at establishing the superiority of custom CNN over pre trained architectures. Further augmentation techniques are superior to imputation techniques in case of missing values.

The dataset used here is a kaggle dataset and consists of null values.

Null vs Non Null

Imputation techniques

Two types of imputations have been implemented here
- Forward fill imputation
- KNN imputation

Augmentation

The following augmentations have been used in our model building.

Augmentation

Architectures

The Custom model and pre trained (NasNetmobile/MobilenetV2) architectures are shown below.

Custom tuned and Keras tuned CNN	Pre trained architectures

RMSE - Custom Models vs Pre Trained models

Custom Models	Pre trained models

It is seen that custom models with augmentation fare better than imputation techniques with the least rmse scores. Similarly Fine tuned MobileNetV2 with augmentation has the least RMSE score among transfer learning models and also among Custom models suggesting that augmentation is a more robust technique than just imputation.

Inference Time / Model parameter

- Manually tuned custom model has lower inference time on CPU than Keras tuned model
- MobilenetV2 yielded lower RMSE than Manually tuned custom model inspite of increased parameters. This can be attributed to the architectural efficiency of the mobilenet architectures.

Model size	Latency