Deriving Rectangular Regions Bounding Box from Overlapped Image Segments Using Labeled Intersecting Points

Numerous neural network models are capable of generating image segments; however, they tend to merge nearby objects or objects situated behind others into a single large region, encompassing multiple overlapped segments. In this paper, we introduce a novel approach for generating individual rectangular overlapped image segments from a larger region, which is produced by applying morphological operations on an image. The proposed algorithm addresses this issue by utilizing labeled intersecting points, enabling the extraction of distinct overlapping segments. To evaluate its effectiveness, we constructed a new dataset specific to this problem domain. The performance of our algorithm is compared with a standard watershed algorithm on this dataset, and our method outperforms with an impressive recall of 96%, F1-score of 97%, and an average precision of 96.21%.