Research map / Point clouds / Segmentation and detection
Generalized Zero-Shot Learning for Point Cloud Segmentation with Evidence-Based Dynamic Calibration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes E3DPC-GZSL, an evidence-based uncertainty-driven dynamic calibrated stacking method with semantic tuning of text embeddings for inductive generalized zero-shot 3D point cloud segmentation on ScanNet v2 and S3DIS.
Problem
Generalized zero-shot semantic segmentation of 3D point clouds suffers from biased predictions toward seen classes, and existing calibrated stacking relies on hyperparameters that must be fixed across all inputs.
Approach
Replace the fixed calibration factor in calibrated stacking with a per-point dynamic factor derived from evidential (Dirichlet-based) uncertainty estimated by a learned module, and refine text embeddings with a learnable scene-conditioned tuning vector before feature synthesis by a GMMN decoder.
Key results
- On ScanNet v2, E3DPC-GZSL reaches 32.0 mIoU (All) and 21.6 HmIoU, versus 30.4 and 20.2 for 3DPC-GZSL.
- On S3DIS, E3DPC-GZSL reaches 50.7 mIoU (All) and 20.4 HmIoU, versus 43.8 and 16.7 for 3DPC-GZSL, a 3.7-point HmIoU gain.
- Ablation shows dynamic calibration adds gains independent of semantic tuning; on S3DIS HmIoU rises from 18.90 (baseline) to 20.42 with all components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace 3DPC-GZSL with E3DPC-GZSL | 3D point cloud GZSL segmentation · ScanNet v2 | HmIoU: +1.4% HmIoU | Helped |
| replace 3DPC-GZSL with E3DPC-GZSL | 3D point cloud GZSL segmentation · S3DIS | HmIoU: +3.7% HmIoU | Helped |
| add dynamic calibration (ablation) | B baseline · ScanNet v2 | HmIoU: 20.77 -> 20.85 | Helped |
| add dynamic calibration (ablation) | B baseline · S3DIS | HmIoU: 17.60 -> 19.28 | Helped |
| add semantic tuning (ablation) | B baseline · ScanNet v2 | HmIoU: 20.77 -> 21.42 | Helped |
| add semantic tuning (ablation) | B baseline · S3DIS | HmIoU: 17.60 -> 18.90 | No clear effect |
| add divergence loss to segmentation loss (ablation) | uncertainty estimator · ScanNet v2 | HmIoU: 21.51 -> 21.53 | No clear effect |
| add divergence loss to segmentation loss (ablation) | uncertainty estimator · S3DIS | HmIoU: 19.17 -> 19.89 | Helped |
| add binary loss to segmentation loss (ablation) | uncertainty estimator · ScanNet v2 | HmIoU: 21.51 -> 21.51 | No clear effect |
| add binary loss to segmentation loss (ablation) | uncertainty estimator · S3DIS | HmIoU: 19.17 -> 19.46 | Helped |
| replace scene-conditioned tuning vector with a random vector (ablation) | semantic tuning · ScanNet v2 | HmIoU: 21.17 ± 0.82 -> 7.83 ± 1.75 | Hurt |
| replace scene-conditioned tuning vector with a random vector (ablation) | semantic tuning · S3DIS | HmIoU: 17.38 ± 3.29 -> 9.98 ± 5.36 | Hurt |
Surprising findings
- Calibrated stacking gives little gain on ScanNet v2 because the model is already highly overconfident there, so subtracting eta does not shift the probability peak.
- Random vectors in place of the learned semantic tuning vector degrade unseen HmIoU sharply (ScanNet v2 HmIoU 21.17 to 7.83).
Limitations
- Gains are modest on ScanNet v2 because the base model is highly overconfident, limiting the effect of calibrated stacking.
- Final predictions still depend heavily on the classifier; calibration alone does not resolve semantic ambiguity (e.g., desk vs. table, door vs. wall).
- Unseen performance remains low in absolute terms (e.g., 15.4 unseen mIoU on ScanNet v2).
Experimental setup
Datasets: ScanNet v2 S3DIS SemanticKITTI
Benchmarks: ScanNet v2 S3DIS SemanticKITTI
Metrics: mIoU HmIoU
Baselines: 3DGenZ 3DPC-GZSL Full supervision
Builds on: calibrated stacking 3DPC-GZSL 3DGenZ generative moment matching network evidential deep learning
Closest papers by meaning
- Generalized Few-shot 3D Point Cloud Segmentation with Vision-Language Model (CVPR 2025 · similarity 0.90)
- Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.89)
- Geometrically-driven Aggregation for Zero-shot 3D Point Cloud Understanding (CVPR 2024 · similarity 0.88)
- Rethinking Few-shot 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.87)
- PDF: A Probability-Driven Framework for Open World 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.87)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.