CenterMask3D: Are Bounding-Box Based Approaches Suitable for Volumetric Instance Segmentation?

We evaluate CenterMask3D, an anchor-free volumetric instance segmentation method adapted from FCOS3D, for the analysis of industrial Computed Tomography data. The study utilizes a specialized dataset of CT scans featuring boxed shoes, characterized by complex geometries and overlapping instance bounding boxes. To address significant memory constraints and limited training samples, we implement a memory-efficient Residual Squeeze-and-Excitation backbone and an extensive on-thefly data augmentation pipeline. Additionally we improve a lightweight Mask-Head based on channel-wise attention. Our model achieves a mean Intersection over Union (mIoU) per instance of 0.365 and an AP@IoU=0.5 of 0.317. While the results demonstrate that CenterMask3D successfully segments individual volumetric objects, the performance highlights inherent limitations of bounding-box-based approaches in dense volumetric settings. Specifically, the accumulation of errors from noisy box proposals and interpolation artifacts suggests that kernel-based alternatives, such as PanopticFCN or Flood Filling Networks, may offer more robust pathways for high-fidelity volumetric instance segmentation.