AGDM-PiMAE and AFP-MonoDETR: a monocular 3D object detection method based on multimodal pretraining and adaptive feature projection

Monocular 3D object detection is vital for cost-effective environmental perception in autonomous driving, focusing on estimating the 3D properties of objects using only a single RGB image. However, the lack of direct depth data causes significant ambiguity in mapping from 2D to 3D, limiting the accuracy of detection. Multimodal pretraining can alleviate this issue by transferring complementary depth-aware features. However, existing frameworks are generally designed for multiple downstream tasks, which weakens their effectiveness for a single task such as monocular 3D detection. To overcome this challenge, we introduce a more advanced framework named AGDM–PiMAE + AFP–MonoDETR. Initially, we present a multimodal pretraining approach inspired by MAE, specifically designed for monocular 3D object detection. This approach is complemented by a module-freezing strategy, focusing computational resources on depth-semantic learning by deactivating irrelevant parameters. Second, to overcome the randomness of static masking, we design an Attention-Guided Dynamic Masking (AGDM) mechanism that adaptively adjusts mask probabilities based on cross-modal attention, enabling pretrained features to better capture depth-sensitive semantic regions. Finally, we propose an Adaptive Feature Projection (AFP) module to improve feature transfer to downstream detection by modeling nonlinear feature distribution differences and reducing semantic information loss. Extensive experiments on the car category of the KITTI dataset demonstrate that the proposed framework improves 𝐴P3𝐷|𝑅40 by 2.40%, 1.57%, and 1.45% under Easy, Moderate, and Hard settings, respectively, validating its effectiveness for monocular 3D object detection