Research map / Point clouds / Point cloud classification

Invaria: Learning Scale and Density Invariance in Point Clouds via Next-Resolution Prediction

NeurIPS 2026 · Point clouds › Point cloud classification · approach: Local feature aggregation networks · 0 citations (OpenAlex)

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 Invaria, a 3D point cloud encoder trained with next-resolution prediction and receptive field calibration to become robust to density and scale shifts, with a smaller model and fewer input tokens.

Problem

Existing 3D point cloud encoders are highly sensitive to sampling resolution (point density) and object scale, degrading sharply when test-time resolution or scale departs from training, which hinders robust deployment such as robotics.

Approach

Train a U-Net point encoder to predict features of higher-resolution point clouds from sparser inputs in latent space, with a cross-resolution latent alignment loss, and calibrate the internal grid size from nearest-neighbor anchor distances so receptive fields adapt to density and scale; at inference, process sparser inputs asymmetrically.

Key results

• On ScanNet at 3x lower resolution (3τ*), Invaria reaches 73.8 mIoU versus 20.5 for PTv3 and 34.2 for Sonata.
• Invaria with 25.5M parameters avoids the 45% size gap of larger baselines and reports 267 GFLOPs versus 352 for PTv3.
• Under scale reduction by 1/3 on ScanNet, Invaria attains 73.8 mIoU, while PTv3 gets 9.4 and Sonata 44.8.

Surprising findings

• Higher-resolution inputs can degrade accuracy more than lower-resolution inputs for existing models, attributed to a reduced receptive field.
• Resolution augmentation of PTv3 lowers native-resolution mIoU by about 2 points (77.6 to 75.6).

Limitations

• Weak on simple planar classes such as picture, door, and window, which are confused with walls.
• Focal loss for class imbalance did not help performance.
• Evaluated only on ScanNet and nuScenes with semantic segmentation as the main task.

Experimental setup

Datasets: ScanNet nuScenes

Benchmarks: ScanNet nuScenes

Metrics: mIoU mAcc allAcc FLOPs memory parameters

Baselines: Sonata Concerto Utonia Sparse MinkUNet PTv3

Builds on: Point Transformer v3 LitePT Visual Autoregressive modeling Lovász-Softmax loss

Closest papers by meaning

• PointInfinity: Resolution-Invariant Point Diffusion Models (CVPR 2024 · similarity 0.89)
• Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.86)
• Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.86)
• Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.86)
• ObjVariantEnsemble: Advancing Point Cloud LLM Evaluation in Challenging Scenes with Subtly Distinguished Objects (AAAI 2025 · similarity 0.86)
• PointMAC: Meta-Learned Adaptation for Robust Test-Time Point Cloud Completion (NeurIPS 2025 · similarity 0.86)