Datasets for face and object detection in fisheye images

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Object and face detection models often struggle with fisheye lens distortions.
• Existing datasets may not adequately represent the unique challenges of fisheye images.
• There is a need for specialized datasets to improve model performance in fisheye imaging.

Purpose of the Study:

• To introduce two novel fisheye image datasets: VOC-360 and Wider-360.
• To provide resources for training and evaluating object and face detection models on fisheye imagery.
• To facilitate research in computer vision tasks involving wide-angle lens distortions.

Main Methods:

• Generated fisheye images by post-processing regular images from VOC2012 and Wider Face datasets.
• Utilized a Matlab-implemented model for mapping regular images to fisheye projections.
• Created VOC-360 (39,575 images) for object detection, segmentation, and classification.
• Created Wider-360 (63,897 images) specifically for face detection.

Main Results:

• Successfully generated two large-scale fisheye image datasets.
• VOC-360 offers diverse data for object-centric computer vision tasks.
• Wider-360 provides extensive data for robust face detection in fisheye views.

Conclusions:

• The VOC-360 and Wider-360 datasets are valuable resources for advancing fisheye image analysis.
• These datasets will accelerate the development of specialized detectors and segmentation modules.
• They serve as a crucial step towards improving AI performance in applications using fisheye cameras.