- Previously, proposed anchor based methods are not scale-invariant hence, unable to detect small objects.
- The paper proposes a scale invariant face detector that resolved the issue with anchor-based models for smaller faces.
- Anchor-based object detection methods that perform binary classification with class prediction scores and bounding-box regressions as output. Using Hard Negative Mining further boosts performance.(Drawbacks: reduced speed when dealing with smaller faces)
- Object detection with context information has been used in models like CMS-RCNN
- Multi-scale SSD model that serves as an inspiration for the current paper.
- Scale Compensation Anchor Matching : distinct anchor scales that improves the recall rate for smaller faces.
- Scale Equitable Model : varying anchor scales for appropriate face detection , improving the detection rate that usually drops in the previous models when detecting small faces.
- Maxout Background label: reduces the high false positive rate of small faces by predicting N_m scores for the background label and choosing the "Max" as the final scores.
-
The dataset used for the training/testing:
- AFW , 205 images with 473 labeled faces
- FDDB, 5171 faces in 2845 images
- Pascal Faces, 1335 labeled faces in 851 images with large face appearance and pose variations.
- WIDER faces , ~32K images with ~400K annotations of varying scales and poses.
-
The dataset is split into easy, medium and hard subsets.