An Efficient Intersection Over Union Algorithm with Angle Orientation for an Improved 3D Object Detection

An essential portion of computer vision research is the detection of objects in three dimensions (3D) that has been used widely in many applications such as autonomous vehicles and augmented reality, visual security surveillance, and many more. The mismatch between the desired ground truth and the predicted bounding boxes is crucial to the process of 3D object detection. This difference is also known as loss function in the field of 3D object detection. Of late, there has been significant work in the loss function algorithms that led to several versions of improvement. However, the influence of angle orientation between the ground-truth bounding box and the predicted box is yet to be addressed. In 3D object detection, the loss function is usually described as an intersection over union (IoU) algorithm. Therefore, in this work, an efficient IoU (EIoU) algorithm with refinement of angle orientation is presented as an improved version of the IoU algorithm. Evaluation outcomes on the KITTI dataset have shown that the projected technique is able to reduce false detection and correct the angle offset significantly as compared to the standard Mixture-Density-based 3D Object Detection (MD3D) method.