【摘要】 实例分割是一项具有挑战性的任务，不仅需要预测场景中每个对象的语义标签，还需要预测实例标识。为了解决上述问题，HAIS提出了一种基于分层聚合的三维实例分割网络，但是在逐点预测阶段，难以精确预测语义标签和准确预测偏移。为了解决以上问题，引入倒置残差模块以提升语义标签的预测准确性，并添加了方向损失以约束预测偏移矢量的方向，从而进一步提高了实例分割的精度。在公共基准数据集ScanNet v2上对的方法进行了评估，验证了它在实例分割任务中的有效性。更多还原

【Abstract】 Instance segmentation is a challenging task that requires predicting not only the semantic labels of each object in the scene, but also the instance identity. To solve this problem, HAIS proposes a 3D instance segmentation network based on hierarchical aggregation, but it is difficult to accurately predict semantic labels and accurately predict offsets in the point prediction stage. To address this problem, we introduce an inverted residual module to improve the prediction accuracy of semantic labels, and add an orientation loss to constrain the direction of the predicted offset vector, which further improves the accuracy of instance segmentation. We evaluate our method on a public benchmark dataset, ScanNet v2,and verify its effectiveness in the instance segmentation task.更多还原