Two-Phase Approach for Monocular Object Detection and 6-DoF Pose Estimation

We present a two-phase algorithm that first identifies the categories and 2D proposal regions of 3D objects and then estimates the eight corners of cubes bounding the target objects. Given the predicted corners, the six-degrees-of-freedom (6-DoF) poses of the 3D objects are calculated using the conventional perspective-n-point (PnP) algorithm and evaluated with respect to manually annotated corners. In addition, several 3D models with high-quality shapes, texture information, 2D images, and annotations, such as 2D boxes, 3D cuboids, and segmentation masks, are collected. New objects are included while validating the proposed method. Our results are compared qualitatively and quantitatively with those of the baseline model using the publicly accessible LineMOD dataset, additional annotations in the OCCLUSION dataset, and our own custom dataset. While handling single and multiple objects in testing scenes, the proposed method is observed to exhibit clear improvements on both the aforementioned datasets and in real-world examples.