CompetitorFormer：通过竞争策略缓解查询冲突的 3D 实例分割
CompetitorFormer: Mitigating Query Conflicts for 3D Instance Segmentation via Competitive Strategy
针对基于 Transformer 的 3D 实例分割中多个查询同时瞄准同一实例导致掩码碎片化的问题，提出查询竞争层与排序交叉注意力，可复用于任何基于查询的 3D/2D 分割框架以加速收敛并提升精度。
原文依据 we observe that multiple queries often target the same instance simultaneously, leading to fragmented masks for a single object
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论基于 Transformer 的方法最近已成为 3D 实例分割的主流范式。这些方法通常采用多层解码器，将一组可学习查询迭代地精炼为实例掩码预测。然而，我们观察到多个查询常常同时瞄准同一个实例，导致单个对象的掩码碎片化。我们将这一现象定义为查询间竞争，它会减慢收敛速度并限制分割精度。为了解决这个问题，我们提出了 CompetitorFormer，一个为基于 Transformer 的方法设计的新框架。我们的方法通过显式建模查询之间的竞争关系来缓解查询间竞争。具体来说，我们在每个解码器阶段之前引入一个查询竞争层，以构建动态竞争格局，使每个查询能够感知其相对重要性。此外，所提出的相对关系编码和排序交叉注意力模块通过优先考虑主导查询来增强自注意力和交叉注意力。大量实验表明，我们的方法收敛更快，并在 ScanNetV2、ScanNet++V2、ScanNet200 和 S3DIS 数据集上取得了更优的性能。代码可在 https://github.com/DuanchuWang/CompetitorFormer 获取。
英文原文摘要 1296 字符
Transformer-based approaches have recently become the dominant paradigm for 3D instance segmentation. These methods typically employ a multi-layer decoder that iteratively refines a set of learnable queries into instance mask predictions. However, we observe that multiple queries often target the same instance simultaneously, leading to fragmented masks for a single object. We define this phenomenon as inter-query competition, which slows convergence and limits segmentation accuracy. To address this problem, we present CompetitorFormer, a novel framework designed for Transformer-based methods. Our method mitigates inter-query competition by explicitly modeling the competitive relationships among queries. Specifically, we introduce a Query Competition Layer before each decoder stage to construct a dynamic competitive landscape, allowing each query to perceive its relative importance. In addition, the proposed Relative Relationship Encoding and Rank Cross-Attention modules enhance both self-attention and cross-attention by prioritizing dominant queries. Extensive experiments show that our approach converges faster and achieves superior performance on the ScanNetV2, ScanNet++V2, ScanNet200, and S3DIS datasets. Code is available at https://github.com/DuanchuWang/CompetitorFormer.