MonoPSTR: Monocular 3-D Object Detection With Dynamic Position and Scale-Aware Transformer

Transformer-based approaches have demonstrated outstanding performance in monocular 3-D object detection, which involves predicting 3-D attributes from a single 2-D image. These transformer-based methods typically rely on visual and depth representations to identify crucial queries related to objects. However, the feature and location of queries are expected to learn adaptively without any prior knowledge, which often leads to an imprecise location in some complex scenes and a long-time training process. To overcome this limitation, we present MonoPSTR, which employs a dynamic position and scale-aware transformer for monocular 3-D detection. Our approach introduces a dynamically and explicitly position-coded query (DEP-query) and a scale-assisted deformable attention (SDA) module to help the raw query possess valuable spatial and content cues. Specifically, the DEP-query employs explicit position priors of 3-D projection coordinates to enhance the accuracy of query localization, thereby enabling the attention layers in the decoder to avoid noisy background information. The SDA module optimizes the receptive field learning of queries by the size priors of the corresponding 2-D boxes; thus, the queries could acquire high-quality visual features. Both the position and size priors do not require any additional data and are updated in each layer of the decoder to provide long-term assistance. Extensive experiments show that our model outperforms all the existing methods in terms of inference speed, which reaches the impressive 62.5 FPS. What is more, compared with the backbone MonoDETR, our MonoPSTR achieves around two times of training convergence speed and surpasses its precision by over 1.15% on famous KITTI dataset, demonstrating the sufficient practical value. The code is available at: https://github.com/yangfan293/MonoPSTR/tree/master/MonoPSTR.