Pixel-Aligned Recurrent Queries for Multi-View 3D Object Detection
This work addresses 3D object detection from multiple camera views for applications like robotics and AR, representing an incremental improvement over existing transformer-based methods.
The paper tackles multi-view 3D object detection by introducing PARQ, which uses pixel-aligned recurrent queries with transformer architecture, resulting in outperforming prior methods on ScanNet and ARKitScenes datasets with faster learning and detection, robustness to distribution shifts, and adaptability to additional views and inference compute.
We present PARQ - a multi-view 3D object detector with transformer and pixel-aligned recurrent queries. Unlike previous works that use learnable features or only encode 3D point positions as queries in the decoder, PARQ leverages appearance-enhanced queries initialized from reference points in 3D space and updates their 3D location with recurrent cross-attention operations. Incorporating pixel-aligned features and cross attention enables the model to encode the necessary 3D-to-2D correspondences and capture global contextual information of the input images. PARQ outperforms prior best methods on the ScanNet and ARKitScenes datasets, learns and detects faster, is more robust to distribution shifts in reference points, can leverage additional input views without retraining, and can adapt inference compute by changing the number of recurrent iterations.