Enhancing 3D-2D Representations for Convolution Occupancy Networks
作者：
Highlights：
• Proposed a Position-Aware Transformer (PAT) architecture that leverages the Adaptive Multi-Scale Position Encoding (AMSPE) to encode both global and local position information for better point feature representations.
• Proposed a 3D Correlation-Guided Enhancement (CGE) network to alleviate the ambiguous or noisy representations in the 2D feature projection.
• The proposed method outperforms the state-of-the-art methods in both quantitative and qualitative comparisons.
摘要
•Proposed a Position-Aware Transformer (PAT) architecture that leverages the Adaptive Multi-Scale Position Encoding (AMSPE) to encode both global and local position information for better point feature representations.•Proposed a 3D Correlation-Guided Enhancement (CGE) network to alleviate the ambiguous or noisy representations in the 2D feature projection.•The proposed method outperforms the state-of-the-art methods in both quantitative and qualitative comparisons.
论文关键词：Implicit 3D representation,Multi-scale 3D position encoding,3D Correlation-Guided Attentions
论文评审过程：Received 8 February 2022, Revised 16 September 2022, Accepted 4 October 2022, Available online 13 October 2022, Version of Record 19 October 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2022.109097
