The 3-dimension (3D) object detection is an important task in autonomous driving and robot vision. Recently, multi-view 3D object detection has become crucial for better understanding the surroundings. In multi-view 3D object detection, there are two ...
The 3-dimension (3D) object detection is an important task in autonomous driving and robot vision. Recently, multi-view 3D object detection has become crucial for better understanding the surroundings. In multi-view 3D object detection, there are two types of approaches: a Bird's-Eyes-View (BEV)-based and a sparse query-based approaches. The BEV-based method can aggregate spatial information across multiple images. However, the BEV-based method lost the height information since BEV feature is represented in 2-dimension (2D) space. Additionally, this method takes longer training and inference times than the sparse query-based due to the view-transformation between 2D and 3D space.
For these reasons, the sparse query-based approach has gained much attention for its efficiency. This method has been explored to enhance connectivity between 3D sparse queries and surrounded 2-dimension (2D) images. It is important to detect an object that appears across two or more images.
In this thesis, a new sparse query-based method is proposed. The method introduces a 4D query designed to clearly define the purpose of an object query to enhance the connectivity between the difference dimensions. The 4D query is defined that includes all target parameters: center, scale, orientation and velocity. However, it may lead unstable training due to its high-dimensional information. To alleviate the complexity of a 4D query, this thesis proposes a training strategy called 4D query denoising. This strategy aims to enhance the training stability and accelerate convergence. Also, the distance-wise feature sampling, a method that considers the relative size of objects based on distance, is introduced. This approach ensures a precise alignment between the position of sparse query and the key from image features. Finally, the proposed network extracts the 2D guidance to generate an initial position of the query by utilizing an auxiliary 2D detector. With promising queries, the query position can be updated easier than existing works.
The proposed method demonstrates a remarkable performance improvement on the nuScenes dataset. In comparison to the StreamPETR as the latest state-of-the-art (SOTA) approach, we achieve the increase of 0.9% in mean Average Precision (mAP), 0.4% of nuScenes Detection Score (NDS), 0.6% of mean Average Translation Error (mATE), 0.2% of mean Average Scale Error (mASE), and 1.6% of mean Average Attribute Error (mAAE). In addition, we prove the proposed method has a faster convergence at least 2 times than the StreamPETR.