Skip to content
Search

Point Transformer V3

lastmod 2026-10-27created 2026-02-136 edits

Point Transformer V3: Simpler, Faster, Stronger

Abstractors

This paper is not motivated to seek innovation within the attention mechanism. Instead, it focuses on overcoming the existing trade-offs between accuracy and efficiency within the context of point cloud processing, leveraging the power of scale. Drawing inspiration from recent advances in 3D large-scale representation learning, we recognize that model performance is more influenced by scale than by intricate design. Therefore, we present Point Transformer V3 (PTv3), which prioritizes simplicity and efficiency over the accuracy of certain mechanisms that are minor to the overall performance after scaling, such as replacing the precise neighbor search by KNN with an efficient serialized neighbor mapping of point clouds organized with specific patterns. This principle enables significant scaling, expanding the receptive field from 16 to 1024 points while remaining efficient (a 3x increase in processing speed and a 10x improvement in memory efficiency compared with its predecessor, PTv2). PTv3 attains state-of-the-art results on over 20 downstream tasks that span both indoor and outdoor scenarios. Further enhanced with multi-dataset joint training, PTv3 pushes these results to a higher level.

Comprehensive comparison of the PTv3 family and competing models (as of 2025)

Comprehensive comparison of the PTv3 family and competing models (as of 2025)

Model

Published

Type

Params

ScanNet mIoU

NuScenes mIoU

Speed/efficiency

Key innovation

PTv3

CVPR'24

Backbone

46.1M

77.5 (scratch) / 79.4 (pretrain)

80.4

Baseline

Serialize-and-patch, xCPE

PTv3 Extreme

2024

Competition

46.1M+

—

1st on Waymo

—

TTA, ensembling, data augmentation

Sonata

CVPR'25

3D SSL Pretrain

~46M (Base)

79.4+ (fine-tune)

—

Same as PTv3

Encoder-only self-distillation

Concerto

NeurIPS'25

2D-3D SSL

39M / 108M / 208M

80.7

—

Same as PTv3

2D-3D cross-modal joint embedding

Flash3D

CVPR'25

Backbone

PTv3-class

Above PTv3

Above PTv3

2.25x faster, 2.4x memory↓

PSH + FlashAttention integration

LitePT

2025.12

Backbone

12.7M ~ 86M

On par with PTv3

82.2 (S)

3.6x parameters↓, 2x faster

Conv-Attn hybrid, Point-ROPE

Summary

Accuracy SOTA

Concerto (NeurIPS'25) is currently the best, with 80.7% mIoU on ScanNet. It uses PTv3 as the backbone but raises representation quality itself through 2D-3D joint self-supervised learning.

Efficiency SOTA

LitePT cuts parameters 3.6x and doubles speed compared with PTv3, while actually performing better outdoors (NuScenes). Flash3D also shows strong efficiency through GPU-architecture-level optimization.

Practical selection criteria

  • Need real-time processing → LitePT or Flash3D (efficiency edge)
  • Need the highest accuracy → Concerto pretrained PTv3 (accuracy edge)
  • General-purpose backbone → PTv3 (most proven ecosystem, Pointcept codebase)

Documentations

https://arxiv.org/abs/2312.10035
[2312.10035] Point Transformer V3 - Simpler, Faster, Stronger
  • LitePT: Lighter Yet Stronger Point Transformer
  • PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection
  • Combinatorial Approximations for Cluster Deletion: Simpler, Faster, and Better
  • PaLI-3 Vision Language Models: Smaller, Faster, Stronger
  • ResT V2: Simpler, Faster and Stronger

See also

Favorite site

author Zer0 sitemap