Face Detection

Purpose

This model detects faces and outputs bounding boxes.

Architecture

It uses InceptionV2 architecture Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift with FPN Feature Pyramid Networks for Object Detection.

Intended Use

General purpose face detector.

Limitations

False positives with respect to small objects that they are being confused with faces. Heavy occlusion is going to hurt performance, but the model has been trained with partial occlusion.

• Other Versions and Major Differences: This model is Inception V2 model with 512 resolution. There are no other versions

The model is trained on openimages_v4_face_only_crowd_filtered.

Training/Test Data and Taxonomy

• The model is trained on openimages_v4_face_only_crowd_filtered.

• Test data is a combination of internal data and the val split of openimages_v4_all_attr.

Evaluation Results

Average Precision (AP)

• AP50 for OpenImagesV4: 82.8

• AP50 for WiderFace (easy): 85.2

• AP50 for WiderFace (medium): 81.1

• AP50 for WiderFace (hard): 61.6