Research map / Point clouds / Point cloud classification
MICAS: Multi-grained In-Context Adaptive Sampling for 3D Point Cloud Processing
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 MICAS, a multi-grained adaptive sampling framework for in-context learning on point clouds that combines task-adaptive Gumbel-softmax point sampling with query-specific learned prompt ranking, built on PIC.
Problem
In-context learning for multi-task 3D point cloud processing is sensitive both across tasks (inter-task) and across prompts for the same task (intra-task) because conventional sampling such as FPS is task- and prompt-agnostic.
Approach
Replaces FPS with a learnable, task-conditioned Gumbel-softmax point sampler that uses prompt features, and adds a PointNet-based prompt scorer trained with a list-wise ranking loss on performance pseudo-labels to pick the best prompt per query.
Key results
- MICAS applied to PIC-Sep improves part segmentation mIOU from 75.0 to 86.8 and PIC-Cat + MICAS reaches 87.9 mIOU, versus 79.0 for PIC-Cat.
- PIC-Cat + MICAS reduces registration CD average from 14.1 to 9.8 (x1000).
- The paper reports a 4.1% improvement in part segmentation as its headline gain; inference time rises roughly threefold (e.g., 15.6 to 47.1 ms for PIC-Cat).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FPS with task-adaptive point sampling (ablation) | PIC-Cat · ShapeNet In-Context denoising | CD (x1000): 6.0 -> 4.7 | Helped |
| Replace FPS with task-adaptive point sampling (ablation) | PIC-Sep · ShapeNet In-Context reconstruction | CD (x1000): 4.4 -> 4.6 | Hurt |
| Replace FPS with task-adaptive point sampling (ablation) | PIC-Sep · ShapeNet In-Context part segmentation | mIOU: 78.7 -> 86.6 | Helped |
| Add query-specific prompt sampling to FPS (ablation) | PIC-Cat · ShapeNet In-Context registration | CD (x1000): 14.5 -> 10.8 | Helped |
| Add query-specific prompt sampling to FPS (ablation) | PIC-Cat · ShapeNet In-Context part segmentation | mIOU: 79.9 -> 80.2 | No clear effect |
| Add query-specific prompt sampling to task-adaptive point sampling (ablation) | PIC-Cat · ShapeNet In-Context denoising | CD (x1000): 5.7 -> 4.6 | Helped |
| Add query-specific prompt sampling to task-adaptive point sampling (ablation) | PIC-Sep · ShapeNet In-Context denoising | CD (x1000): 6.5 -> 5.1 | Helped |
| Add query-specific prompt sampling to task-adaptive point sampling (ablation) | PIC-Sep · ShapeNet In-Context registration | CD (x1000): 3.5 -> 3.7 | Hurt |
| Add MICAS adaptive point and prompt sampling | PIC-Sep · ShapeNet In-Context registration | CD (x1000): 10.3 -> 3.7 | Helped |
| Add MICAS adaptive point and prompt sampling | PIC-Cat · ShapeNet In-Context part segmentation | mIOU: 79.0 -> 87.9 | Helped |
Surprising findings
- Task-adaptive point sampling is slightly worse than FPS on reconstruction with PIC-Sep (0.2 higher average CD), so it does not uniformly help all tasks.
Limitations
- Inference time increases roughly threefold with the full method.
- Task-adaptive sampling slightly underperforms FPS on reconstruction for PIC-Sep.
- Evaluated only on the ShapeNet In-Context benchmark with four tasks.
Experimental setup
Datasets: ShapeNet In-Context Dataset ShapeNet ShapeNetPart
Benchmarks: ShapeNet In-Context Dataset
Metrics: Chamfer Distance mIOU
Baselines: PIC-Cat PIC-Sep PIC-S-Cat PIC-S-Sep Point-BERT Point-MAE ACT I2P-MAE ReCon PointNet DGCNN PCT
Builds on: PIC Gumbel-softmax UDR PointNet SkeletonNet
Closest papers by meaning
- LTA-PCS: Learnable Task-Agnostic Point Cloud Sampling (CVPR 2024 · similarity 0.89)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.87)
- SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity (CVPR 2025 · similarity 0.87)
- GCD-Sampling: A General Cross-scale Decoupled Sampling for Point Cloud (AAAI 2025 · similarity 0.87)
- Improved MLP Point Cloud Processing with High-Dimensional Positional Encoding (AAAI 2024 · similarity 0.86)
- SuperPC: A Single Diffusion Model for Point Cloud Completion, Upsampling, Denoising, and Colorization (CVPR 2025 · similarity 0.86)
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.