Optimizing Monocular 3D Object Detection on KITTI: Harnessing Power of Right Images

Monocular 3D object detection is an important yet challenging problem in computer vision, with applications such as autonomous driving. A key limitation in advancing this field is the scarcity of annotated training data, an issue exacerbated in benchmarks like KITTI, which provide only around 7,000 labeled images. Prior arts have developed techniques to improve monocular 3D detection but often rely on external sources of data like LiDAR to supplement the limited training images. In this work, we propose a pre-training strategy that addresses the limited data issue by leveraging unlabeled right-camera images available within the KITTI dataset itself. We pre-train a model initialized for 3D detection by using “right” views before fine-tuning on just “left” images. Our experiments validate that, through this strategic pre-training with readily available “right” images, significant improvements can be achieved over models trained from scratch on only “left” images. We observe consistent gains in 3D detection performance when leveraging “right” image pre-training, without requiring any external LiDAR data. Our method provides evidence that mining unlabeled or weakly labeled in-domain data can effectively remedy the pervasive challenge of limited training data for monocular 3D object detection. This offers a plug-and-play practical strategy to use available datasets like KITTI better and reduce overfitting.