Towards Generalizable Distance Estimation By Leveraging Graph Information

Approximating the distance of objects present in an image remains an important problem for computer vision applications. Current SOTA methods rely on formulating this problem to convenience depth estimation at every pixel; however, there are limitations that make such solutions non-generalizable (i.e varying focal length). To address this issue, we propose reformulating distance approximation to a per-object detection problem and leveraging graph information extracted from the image to potentially achieve better generalizability on data acquired at multiple focal lengths.