PODE: Panoptic-Aware Object-Wise Depth Estimation for Occlusion Prediction

Depth estimation of an object detected in a monocular camera view remains a famous challenge. The existing approaches to estimate object-wise depth use additional sensors such as stereo vision or LiDAR sensors to estimate depth dimension. However, some recent research attempts to estimate depth in monocular camera view via deep-learning methods. In this research, we propose a Panoptic-aware Object-wise Depth Estimation (PODE) model that achieves panoptic segmentation along with depth estimation of the object as seen in a monocular camera view. The proposed model, having joint panoptic segmentation and depth estimation neural network based on the feature pyramid network architecture to obtain instance-wise depth values uniquely, identifies every instance of an object class along with its depth dimension. This gives a clear advantage in case of a potential occlusion of the target object under track. The knowledge of object-wise depth estimation can predict occluder and occludee based on their estimated depth even in cases of objects falling in same object class with different instances. The proposed model is tested on some established as well as self-acquired CamPed dataset (https://github.com/RahulRaman2/CamPed-Dataset). The test results of the proposed model are quite motivating and outperform comparatively complex state-of-the-art architectures in this domain.