Computer Vision (CV) has so many applications such as but not limited to object recognition, which is a collection of computer vision tasks that involves identifying objects in images. One of CV applications is People counting, and it is useful for automatically counting the number of persons in a class, or a ceremony, or an event. People counting is based on face detection is a challenging task and still an open problem in computer vision. This research investigates two object detection models for detecting and counting people's faces. The first model is based on Faster-RCNN and the second one is based on SSD. These models are deep neural networks that are trained on object detection tasks. In this work, we train Faster-RCNN and SSD models on Wider-Face dataset, which is composed of faces in a variety of conditions relating to occlusion, illumination, expression, pose and scale. The evaluation result on the test part of the wider face dataset is 0.5 of accuracy for Faster-RCNN and SSD, also the Mean Relative Error for the Faster-RCNN is 0.3 and the SSD is 0.4. The Mean Absolute Error for the Faster-RCNN is 7.5 and the SSD is 8.6.
Detecting and Counting People's Faces in Images Using Convolutional Neural Networks
Saad M.
Secondo
Supervision
;
2021-01-01
Abstract
Computer Vision (CV) has so many applications such as but not limited to object recognition, which is a collection of computer vision tasks that involves identifying objects in images. One of CV applications is People counting, and it is useful for automatically counting the number of persons in a class, or a ceremony, or an event. People counting is based on face detection is a challenging task and still an open problem in computer vision. This research investigates two object detection models for detecting and counting people's faces. The first model is based on Faster-RCNN and the second one is based on SSD. These models are deep neural networks that are trained on object detection tasks. In this work, we train Faster-RCNN and SSD models on Wider-Face dataset, which is composed of faces in a variety of conditions relating to occlusion, illumination, expression, pose and scale. The evaluation result on the test part of the wider face dataset is 0.5 of accuracy for Faster-RCNN and SSD, also the Mean Relative Error for the Faster-RCNN is 0.3 and the SSD is 0.4. The Mean Absolute Error for the Faster-RCNN is 7.5 and the SSD is 8.6.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
