Deep-Learning Based Object Detection and Shape Recognition in Multiple Occluded Images

The 2D to 3D reconstruction process suffers from quality issues such as illuminance, texture, noise, color, and occlusions. This work focuses on the issue of occlusion between multiple objects either full or partial overlapping with the same or different objects. The intra-object occlusion among the objects in the single RBG image is the input for accurate object detection using the novel context and object shape recognition (COSR) approach for good-quality reconstruction. The proposed occlusion detection mechanism intends to recognize the shape of the objects among the more minor and more significant things. The occlusion patterns are varied for different object categories, to enable object detection with the bounding box algorithm which exploits the reference objects. Object shapes are represented and distinguished among multiple objects using a slicing operation based on the object’s features like area, height, and width. The proposed feature-based model of COSR used different data representation methods to provide realism in the reconstruction process for the improved plausibility appearance among the multiple objects in partial occlusions.