TopFormer: Topology-Aware Transformer for Point Cloud Registration

Point cloud registration requires finding reliable correspondences between two scans of the same scene. Transformer-based methods have improved on classical descriptor matching by using attention to capture global context, yet they remain vulnerable to scenes with repetitive geometric motifs—tiled floors, uniform walls, symmetric objects—where many local regions appear nearly identical. TopFormer tackles this robustness gap by enriching transformer feature tokens with topological structure descriptors that capture how each surface point connects to its neighbours at multiple scales. These topological features provide information not available in standard geometric or learned local descriptors: two patches that look locally similar may have different global connectivity signatures, making them distinguishable. TopFormer integrates these topology-aware tokens into a standard transformer matching pipeline with minimal added complexity. Presented at the International Conference on Computational Visual Media (CVM) 2024, the method achieves competitive or superior registration recall on 3DMatch and KITTI benchmarks, with the clearest gains on the challenging low-overlap evaluation split where conventional attention-based methods degrade most.

Problem setting

Transformer-based point cloud registration methods have achieved strong performance by matching local geometric descriptors with global attention, but they still struggle in scenes with repetitive geometry or limited texture where local shape alone is insufficient to distinguish correspondences. TopFormer addresses this by incorporating topological structure descriptors—derived from persistent homology or graph connectivity of the local neighbourhood—into the transformer feature tokens. These topology-aware tokens encode how each point sits within the broader surface topology, providing a complementary discriminative signal that reduces false matches in symmetric or repetitive regions.

In the broader publication record, this work appears in International Conference on Computational Visual Media (CVM) 2024, 112–128. The visual notes below pair the paper’s original figures with a concise reading of the method, experimental setup, and reported results.

Method and visual evidence

The method works on 3D geometric observations such as point clouds, poses, correspondences, or segmented regions, then uses the proposed representation to improve robustness under noise, viewpoint change, or limited observations.

The extracted figures below show the geometric representation, network or optimization pipeline, and qualitative or quantitative results.

Method overview. This image is extracted from an embedded PDF image object on page 2, then recomposed for web display.

Representation and setup. This image is extracted from an embedded PDF image object on page 2, then recomposed for web display.

Experimental evidence. This image is extracted from an embedded PDF image object on page 5, then recomposed for web display.

Result comparison. This image is extracted from an embedded PDF image object on page 5, then recomposed for web display.

Additional visual result. This image is extracted from an embedded PDF image object on page 11, then recomposed for web display.

Results and impact

The evaluation reported in International Conference on Computational Visual Media (CVM) 2024, 112–128 uses the extracted figures above to show the method’s measurement, reconstruction, segmentation, matching, or diagnostic behavior on representative experiments. These visuals are paired with the paper’s quantitative or qualitative analysis to make the workflow easier to inspect from the homepage.

Source handling

I extracted 24 candidate image objects from paper.pdf and generated the compressed WebP figures used on this page.