Abstract: The current 3D plane recovery method can not effectively restore the target plane of small objects in the scene. To solve this problem, a 3D plane recovery method based on visual attention mechanism is proposed. Specifically, low-scale features were down-sampled and fused with high-scale features. An adaptive multi-head attention mechanism was utilized to encode the fused features. A plane segmentation module and a plane parameter estimation module were defined, and the convolutional neural network and the visual attention mechanism decoded the features. Extensive experimental results show that the proposed method can effectively restore the object plane representation of small objects, and it is significantly competitive with the current state-of-the-art 3D plane recovery methods.