# Boosting Monocular 3D Object Detection with Object-Centric Auxiliary Depth Supervision

**Authors:** Youngseok Kim, Sanmin Kim, Sangmin Sim, Jun Won Choi, Dongsuk Kum  
**arXiv:** [2210.16574](https://arxiv.org/abs/2210.16574)  
**Date:** 2022-10-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.16574v1/

## Key points

- Authors present a novel approach to enhancing monocular 3D object detection
- Proposed method improves RGB image-based 3D detection by training the network with depth prediction loss
- Object-centric depth prediction loss focuses on foreground objects for more accurate depth estimation
- Depth regression model predicts uncertainties in depth values, providing insights into confidence levels of detected objects
- Tailored network architecture designed for end-to-end training with raw LiDAR points effectively
- Extensive experiments show significant improvement in monocular image-based 3D detectors while maintaining real-time inference speeds

## Summary

In their paper titled "Boosting Monocular 3D Object Detection with Object-Centric Auxiliary Depth Supervision," authors Youngseok Kim, Sanmin Kim, Sangmin Sim, Jun Won Choi, and Dongsuk Kum present a novel approach to enhancing monocular 3D object detection. Recent advancements in this field have utilized a depth estimation network as an intermediary step in the 3D detection process. While depth map techniques have shown superior accuracy in determining object depths compared to other methods, they can be limited by the precision of the depth map itself. Additionally, using separate networks for depth estimation and 3D detection can significantly increase computational costs and inference times. To address these challenges, the authors propose a method that improves RGB image-based 3D detection by training the detection network with a depth prediction loss similar to the depth estimation task. This allows for more robust supervision from raw LiDAR points without requiring additional human annotation costs. By focusing on foreground objects through an object-centric depth prediction loss, which leverages pixel-wise supervision in an object-specific manner, the proposed approach enhances accuracy in estimating object depths crucial for effective 3D object detection. Moreover, the authors introduce a depth regression model trained to predict uncertainties in depth values, providing insights into the confidence levels of detected objects in three dimensions. To facilitate end-to-end training with raw LiDAR points effectively, they reevaluate regression targets for 3D objects and design a tailored network architecture. Extensive experiments conducted on benchmark datasets such as KITTI and nuScenes demonstrate that their method significantly boosts monocular image-based 3D detectors beyond traditional depth map approaches while maintaining real-time inference speeds. The research contributes valuable insights into advancing monocular 3D object detection through innovative approaches to leveraging auxiliary depth supervision and enhancing overall system performance.