PointTPA: Dynamic Network Parameter Adaptation for 3D Scene Understanding

Siyuan Liu; Chaoqun Zheng; Xin Zhou; Tianrui Feng; Dingkang Liang; Xiang Bai

cs.CVApr 6, 2026

PointTPA: Dynamic Network Parameter Adaptation for 3D Scene Understanding

Authors: Siyuan Liu, Chaoqun Zheng, Xin Zhou, Tianrui Feng, Dingkang Liang, Xiang Bai

Organizations: Huazhong University of Science and Technology

Abstract

Scene-level point cloud understanding remains challenging due to diverse geometries, imbalanced category distributions, and highly varied spatial layouts. Existing methods improve object-level performance but rely on static network parameters during inference, limiting their adaptability to dynamic scene data. We propose PointTPA, a Test-time Parameter Adaptation framework that generates input-aware network parameters for scene-level point clouds. PointTPA adopts a Serialization-based Neighborhood Grouping (SNG) to form locally coherent patches and a Dynamic Parameter Projector (DPP) to produce patch-wise adaptive weights, enabling the backbone to adjust its behavior according to scene-specific variations while maintaining a low parameter overhead. Integrated into the PTv3 structure, PointTPA demonstrates strong parameter efficiency by introducing two lightweight modules of less than 2% of the backbone's parameters. Despite this minimal parameter overhead, PointTPA achieves 78.4% mIoU on ScanNet validation, surpassing existing parameter-efficient fine-tuning (PEFT) methods across multiple benchmarks, highlighting the efficacy of our test-time dynamic network parameter adaptation mechanism in enhancing 3D scene understanding. The code is available at https://github.com/H-EmbodVis/PointTPA.

Figures & tables

Figure 1: (a) Scene-level point clouds have more points and highly imbalanced category distributions compared to object-level point clouds. (b) Different from the existing PEFT methods, our PointTPA provides dynamic projection weights during inference and performs better in 3D semantic segmentation.

Figure 2: Overview of PointTPA. It consists of a Serialization-based Neighborhood Grouping (SNG) and a Dynamic Parameter Projector (DPP). During the training process, we freeze the backbone parameter and only fine-tune our DPP and static Adapter modules. The DPP can produce dynamic weights during inference. The same color denotes tokens with spatially close positions.

Figure 3: Illustration of our mixed-insertion strategy. PointTPA is applied to the last block of each stage, while static adapters are adopted in the remaining blocks.

	Reference	Param. (M)	ScanNet Val [ 8 ]			S3DIS Area5 [ 3 ]			ScanNet++ Val [ 54 ]			Avg. Rank
	Reference	Param. (M)	mAcc	allAcc	mIoU	mAcc	allAcc	mIoU	mAcc	allAcc	mIoU	Avg. Rank
Training from scratch
SparseUNet [ 6 ]	CVPR 19	39.2	80.2	90.0	72.3	72.5	89.8	66.3	38.4	80.1	28.8	-
PointNeXt-XL [ 38 ]	NeurIPS 22	41.6	-	-	71.5	-	90.6	70.5	-	-	-	-
PTv3 [ 47 ]	CVPR 24	124.8	85.0	92.0	77.6	78.9	91.7	73.4	53.4	85.6	42.1	-
Full Fine-tuning

Table 1: Results on ScanNet Validation [ 8 ] , S3DIS [ 3 ] , and ScanNet++ [ 54 ] . Only trainable parameters are reported. The dec. and w/o dec. indicate Sonata with and without the decoder, respectively. We use linear probing as the baseline and Sonata (w/o dec.) as a strong comparison. The Avg. Rank is the average of all the metrics of these benchmarks.

Figure 4: Comparison of PEFT methods and FFT on ScanNet [ 8 ] in terms of average training and inference time per sample.

	Param. (M)	ScanNet Val [ 8 ]
		mAcc	allAcc	mIoU
Full Fine-tuning
PTv3-PPT [ 49 ]	124.8	86.0	92.5	78.6
Sonata ( dec. ) [ 46 ]	124.8	86.1	92.5	79.4
General PEFT methods
Sonata (decoder probing)	16.3 (13.1%)	85.8	92.4	78.6

Table 2: Results on ScanNet Validation [ 8 ] with decoder.

Table 3: Ablation on the proposed modules, the number of dynamic bases in the parameter base set (Sec. 3.1 ), and the number of groups in the first encoder stage. “1*” denotes no grouping.

Table 4: Ablation on different settings. We evaluate the effects of (a) scale, (b) DPLayer position, and (c) different space-filling curves in SNG. “w/o” denotes the DPP without the DPLayer. Up and Down refers to applying DPLayer to the up and down projectors.

Inserting Strategy	mIoU	mAcc	allAcc
Dense	77.8	86.6	92.1
Every two blocks	77.9	85.8	92.1
Every three blocks	77.7	85.7	92.1
Last block	77.6	85.4	91.2
First block in every stage	78.0	85.8	92.1
Last block in every stage	78.4	86.3	92.3

Table 5: Ablation on the insertion strategy of our DPP. “Dense” denotes applying DPP in every block.

Figure 5: Qualitative comparison against IDPT [ 61 ] . Green and red boxes indicate correct and incorrect segmentations, respectively, with GT denoting the ground truth.

Figure 6: Visualization of the similarity of dynamic weights.

Table 1: Ablation on ranks and grouping strategies. Here, N denotes the number of points per group.

Method	Train (ms)	Infer (ms)
SNG	+9	+2
DPP	+10	+0
Both	+22	+2

Table 13

	Param. (M)	Reference	ScanNet200 Val [ 39 ]
			allAcc	mIoU
Full Fine-tuning
Sonata( w/o dec. ) [ 46 ]	108.5	CVPR 25	83.8	33.9
General PEFT methods
Sonata( linear probing )	0.2 (0.2%)	CVPR 25	81.1	28.4
+ Prefix Tuning [ 25 ]	1.2 (1.1%)	ACL 21	81.5( +0.4 )	29.4( +1.0 )

Table 2: Validation results on ScanNet200 [ 39 ] .

Figure 1: A comparison of FFT, IDPT [ 61 ] , DAPT [ 70 ] , and PointTPA on segmentation performance, evaluated on ScanNet [ 8 ] .

Figure 2: More visualizations of the semantic segmentation results of our PointTPA on four large-scale scene datasets. (a) ScanNet [ 8 ] , (b) ScanNet200 [ 39 ] , (c) ScanNet++ [ 54 ] , (d) S3DIS [ 3 ] with 3 views.

Method	Reference	mIoU
Sonata (w/o dec.)	CVPR 25	75.9
Sonata (lin)	CVPR 25	70.0
+ DAPT	CVPR 24	74.1( +4.1 )
+ PointTPA (ours)	–	75.8 ( +5.8 )

Table 3: Test on ScanNet [ 8 ]

3D Scene Understanding

3D scene understanding builds geometric and semantic representations of environments for spatial reasoning, perception, and interaction.

See all 293 papers

Partition-Invariant Tuning for 3D Scene Understanding
Sep 11, 2026Hongqiang Lin, Tianle Wang, Shuiwang Li +4Point Cloud Learning Parameter-Efficient Fine-Tuning
ZJU · GUT · THU +3
Point Ladder Tuning: Parameter-Efficient Hierarchical Adaptation for 3D Point Cloud Understanding
Jul 21, 2026Junlin Chang, Longhao Zou, Rui Li3D Point Cloud Segmentation Point Cloud Classification
Beihang University, Beijing · Pengcheng Laboratory, Shenzhen
Position Anchor Tuning: Towards Efficient Adaptation of Pre-Trained Point Cloud Transformers
Sep 16, 2026Zheng Liu, Xin Gao, Jinchao Zhu +1Parameter-Efficient Fine-Tuning Efficient Transformer Inference
School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, China, and also with the Beijing Engineering Research Center of Industrial Spectrum Imaging, Beijing 100083, China · College of Software, Nankai University, Tianjin 300350, China · Department of Automation, BNRist, Tsinghua University, Beijing 100084, China

Sep 11, 2026J/K move · Enter open · S save

Partition-Invariant Tuning for 3D Scene Understanding

Hongqiang Lin, Tianle Wang, Shuiwang Li, Dongxu Zhang+3

ZJU · GUT · THU+3

Point Cloud Learning Parameter-Efficient Fine-Tuning 3D Scene Understanding

Open paper AskPDF arXiv

Scene-level point cloud understanding remains challenging due to diverse geometries and spatial layouts. While pre-trained 3D point cloud foundation models (PFMs) offer strong transferability, full fine-tuning (FFT) incurs substantial computational and storage costs. Parameter-efficient fine-tuning (PEFT) provides a promising alternative, but existing PEFT methods largely focus on object-level point clouds and overlook serialization-induced partition variations in large-scale scenes. To address this issue, we propose PointPiT, a partition-invariant tuning framework for scene-level point clouds. Specifically, a Scene-aware Structural Adapter (SSA) integrates local geometric patterns with global scene context to mitigate partition-induced representation shifts. Moreover, Gradient Subspace Optimization (GSO) selects informative and partition-stable update directions, suppressing partition-dependent variations during optimization. Extensive experiments across multiple scene-level benchmarks demonstrate that PointPiT achieves competitive or even superior performance to full fine-tuning with less than 1% of backbone's parameters, while achieving consistent state-of-the-art performance among representative PEFT methods.

Abstract

3D Scene Understanding

Explore similar work

Partition-Invariant Tuning for 3D Scene Understanding

Point Ladder Tuning: Parameter-Efficient Hierarchical Adaptation for 3D Point Cloud Understanding

Position Anchor Tuning: Towards Efficient Adaptation of Pre-Trained Point Cloud Transformers