Pixel-Aligned Recurrent Queries for Multi-View 3D Object Detection

TL;DR

PARQ introduces a novel multi-view 3D object detection method using transformer-based pixel-aligned recurrent queries, improving accuracy, robustness, and efficiency over previous approaches by leveraging appearance-enhanced queries and recurrent cross-attention.

Contribution

The paper proposes PARQ, a transformer-based 3D detector that uses pixel-aligned recurrent queries initialized from reference points, enabling better 3D-2D correspondence and contextual encoding.

Findings

Outperforms previous methods on ScanNet and ARKitScenes datasets.

Learns and detects faster than prior approaches.

More robust to distribution shifts and can leverage additional views without retraining.

Abstract

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.