FFNet: Frequency Fusion Network for Semantic Scene Completion

Semantic scene completion (SSC) requires the estimation of 3D geometric occupancies objects in scene, along with object categories. Currently, many methods employ RGB-D images to capture and semantic information objects. These use simple but popular spatial- channel-wise operations, which fuse RGB depth data. Yet, they ignore large discrepancy data uncertainty measurements To solve this problem, we propose Frequency Fusion Network (FFNet), a novel method for boosting by better utilizing FFNet explicitly correlates frequency domain, different from features directly extracted convolution operation. Then, network uses correlated guide feature learning RG- B images, respectively. Moreover, accounts properties components RGB- D features. It has learnable elliptical mask decompose learned attending various frequencies facilitate correlation process We evaluate intensively on public SSC benchmarks, where surpasses state-of- the-art methods. The code package is available at https://github.com/alanWXZ/FFNet.