Self-Supervised 3D Monocular Object Detection by Recycling Bounding Boxes

TL;DR

This paper introduces a self-supervised learning pretext task for monocular 3D object detection that recycles bounding boxes, leading to improved detection accuracy especially for low frequency classes.

Contribution

It pioneers the use of bounding box recycling as a self-supervised pretext task in monocular 3D detection, enhancing performance on imbalanced datasets.

Findings

2-3% improvement in mAP 3D and BEV scores with SSL

4-5% increase in ICFW metric for low frequency classes

Effective handling of foreground-background imbalance

Abstract

Modern object detection architectures are moving towards employing self-supervised learning (SSL) to improve performance detection with related pretext tasks. Pretext tasks for monocular 3D object detection have not yet been explored yet in literature. The paper studies the application of established self-supervised bounding box recycling by labeling random windows as the pretext task. The classifier head of the 3D detector is trained to classify random windows containing different proportions of the ground truth objects, thus handling the foreground-background imbalance. We evaluate the pretext task using the RTM3D detection model as baseline, with and without the application of data augmentation. We demonstrate improvements of between 2-3 % in mAP 3D and 0.9-1.5 % BEV scores using SSL over the baseline scores. We propose the inverse class frequency re-weighted (ICFW) mAP score that highlights improvements in detection for low frequency classes in a class imbalanced dataset with long tails. We demonstrate improvements in ICFW both mAP 3D and BEV scores to take into account the class imbalance in the KITTI validation dataset. We see 4-5 % increase in ICFW metric with the pretext task.