ENiqab-V2: A five-point facial landmark dataset for niqab-occluded faces
Description
This article describes ENiqab-V2, an image dataset for facial landmark localization on faces subject to heavy niqab occlusion. The dataset contains 40,533 images and 44,802 annotated faces. Every face carries a bounding box and five facial landmarks left eye, right eye, nose, left mouth corner and right mouth corner stored in the YOLO pose label format at twenty values per face. Just over two thirds of the annotated faces are occluded by eighty percent or more, a regime under which the nose and both mouth corners are not directly observable. The dataset is assembled from five documented sources. A corpus of 19,114 photographs of niqab-wearing individuals, manually annotated in Label Studio, forms the real-photograph foundation. This is extended by 13,673 programmatic zoom crops, 5,783 synthetic-occlusion images in which a code drawn niqab, surgical mask or scarf is warped onto a lightly-occluded face while the original landmark coordinates are preserved unchanged, a 1,363-image clear-face supplement drawn from the WIDER FACE benchmark, and 600 auto-labelled clear-face images. All generated and augmented images are confined to the training partition; the validation and test partitions contain only real, manually annotated photographs, so held-out evaluation is not contaminated by generated data. Annotations were refined by an automated correction pass driven by two independent third-party face detectors, applying 1,918 keypoint corrections, 1,558 additions and 3 relocations, followed by a two-stage quality removal of 116 images containing no usable face. Every count reported here was verified by two independent enumerations of the released files. The dataset supports the development and benchmarking of landmark localization methods in an occlusion regime that existing public benchmarks do not cover: whereas standard occlusion datasets contain partial or incidental occlusion, the faces here are predominantly covered by a contiguous garment. It is distributed in a ready-to-train directory layout with a dataset configuration file, a per-image provenance manifest, and a dataset card recording the exact counts.