SPFormer: Superpoint Pooling and Query-Based Instance Segmentation

Summary

The article introduces Superpoint Transformer (SPFormer), an end-to-end method for segmenting individual objects in sparse 3D point clouds. A sparse 3D U-Net first extracts point features, which are pooled into superpoints to reduce the representation size. Learnable query vectors then attend to those superpoints in a Transformer decoder and predict object classes, confidence, and masks. Mask-informed attention focuses later decoding on relevant regions; bipartite matching trains predictions directly, and inference does not require non-maximum suppression.

The practical section describes an MQL5 adaptation, including custom masked-attention kernels, and presents trading as a possible application to market-data segmentation and signal prediction. The article says its models were trained and tested on historical data and reports profitable results, but gives no performance figures or detailed evaluation in the supplied text. The proposed transfer from 3D scene segmentation to market data remains an interpretation, and the author calls for longer training and thorough validation before live deployment.

Key ideas

• SPFormer compresses point features into superpoints before applying query-based instance prediction.
• Learnable queries use cross-attention to superpoints and predict classes, confidence scores, and instance masks.
• Mask-derived attention restricts later decoding to superpoints likely to belong to a foreground instance.
• The model uses bipartite matching for end-to-end training and avoids non-maximum suppression at inference.
• The trading adaptation reports historical testing but does not provide enough detail to assess robustness or live performance.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.