Organizations: the School of Artificial Intelligence and Robotics, Quanzhou Institute of Industrial Design and Machine Intelligence Innovation, Hunan University, Changsha 410082, China. · the College of Engineering and Design, Hunan Normal University, Changsha 410081, China. · the College of Intelligence Science and Technology, National University of Defense Technology, Changsha 410073, China. · the School of Computing and Information Systems, The University of Melbourne, 3052 Victoria, Australia.
Multi-Layer Perceptron (MLP) models are the foundation of contemporary point cloud processing. However, their complex network architectures obscure the source of their strength and limit the application of these models. In this article, we develop a two-stage abstraction and refinement (ABS-REF) view for modular feature extraction in point cloud processing. This view elucidates that whereas the early models focused on ABS stages, the more recent techniques devise sophisticated REF stages to attain performance advantages. Then, we propose a High-dimensional Positional Encoding (HPE) module to explicitly utilize intrinsic positional information, extending the ``positional encoding'' concept from Transformer literature. HPE can be readily deployed in MLP-based architectures and is compatible with transformer-based methods. Within our ABS-REF view, we rethink local aggregation in MLP-based methods and propose replacing time-consuming local MLP operations, which are used to capture local relationships among neighbors. Instead, we use non-local MLPs for efficient non-local information updates, combined with the proposed HPE for effective local information representation. We leverage our modules to develop HPENets, a suite of MLP networks that follow the ABS-REF paradigm, incorporating a scalable HPE-based REF stage. Extensive experiments on seven public datasets across four different tasks show that HPENets deliver a strong balance between efficiency and effectiveness. Notably, HPENet surpasses PointNeXt, a strong MLP-based counterpart, by 1.1% mAcc, 4.0% mIoU, 1.8% mIoU, and 0.2% Cls. mIoU, with only 50.0%, 21.5%, 23.1%, 44.4% of FLOPs on ScanObjectNN, S3DIS, ScanNet, and ShapeNetPart, respectively. Source code is available at https://github.com/zouyanmei/HPENet_v2.git.
Figures & tables
Fig. 1: Segmentation performance of HPENet V2 on S3DIS [ 4 ] . Bubble diameter is proportional to model parameter and the legend lists reference sizes of 1M, 3M, 20M, and 40M. HPENet V2 achieves higher throughput and competitive or better mIoU than state-of-the-art methods, including PointMetaBase [ 5 ] , PointNeXt [ 6 ] , and PointVector [ 7 ] , while using fewer parameters. Compared with our previous HPENet [ 8 ] , HPENet V2 delivers comparable performance while using fewer parameters and achieving about 2.2× faster inference.
Fig. 2: HPENet V2 architecture for semantic segmentation. The network delineates between the Abstraction (ABS) and Refinement (REF) stages of feature extraction and employs the proposed High-dimensional Positional Encoding (HPE) module in both stages. The differences between HPENet V2 and HPENet [ 8 ] are highlighted with red borders, such as the Backward Fusion Module (BFM) and the non-local MLPs.
Fig. 3: Illustration of abstraction and refinement (ABS-REF) perspective. Left: The proposed ABS-REF view of point cloud models is analogous to subsampling and convolution block view in image models. The shown ABS-REF column expands the abstraction and refinement stages. Right: Representative instantiations of the ABS-REF framework. Whereas early methods, e.g., PointNet++ [ 13 ] , PointConv [ 14 ] , ignore the REF stage, more recent techniques, e.g., Point Transformer [ 11 ] and PointMixer [ 19 ] , achieve higher performance by accounting for the REF stage in point cloud models. Abbreviations include SOP: Symmetric OPeration, OP: aggregation OPeration, PT: Point Transformer, and HPE: proposed High-dimensional Positional Encoding.
Fig. 4: Illustration of separable Abstraction and Refinement (ABS-REF) perspective. Left: The separable ABS-REF view also belongs to the proposed ABS-REF by disentangling the intra-set operation and inter-set operation into non-local MLPs and the reduction operation. Right: Representative instantiations of the separable ABS-REF framework and the evolution of HPENets, i.e., from HPENet to HPENet V2. Please refer to Fig. 3 for the feature propagation in the proposed ABS-REF view.
Fig. 5: Illustration of local aggregation in the ABS stages of MLP-based methods. (a) PreConv processes input point sets with non-local MLPs before sampling operations. (b) Traditional Conv processes neighboring point sets generated by grouping operations with local MLPs. (c) ProConv processes sampled point sets with non-local aggregation after reduction operations. Please refer to the text for further notations.
Line
ABS
REF
OA (%)
mAcc (%)
mIoU (%)
△
Param. (M)
FLOPs (G)
TP (ins./sec.)
1
Conv
Conv
89.0 ± 0.1
73.0 ± 0.4
66.6 ± 0.4
-
1.58
6.93
317
2
Conv*
Conv*
86.3 ± 0.3
67.4 ± 0.6
60.5 ± 0.6
-6.1
1.58
6.76
337
3
PreConv
Conv
88.3 ± 0.4
71.9 ± 0.3
65.6 ± 0.4
-1.0
1.58
5.09
355
4
ProConv
Conv
88.1 ± 0.1
71.7 ± 0.5
65.1 ± 0.4
-1.5
1.58
4.90
380
5
Conv
PreConv
88.6 ± 0.3
72.1 ± 0.4
65.8 ± 0.6
-0.8
1.58
3.01
388
6
Conv
ProConv
87.9 ± 0.1
70.4 ± 0.4
63.7 ± 0.4
-2.9
1.58
3.01
388
TABLE I: A study of local aggregation on S3DIS Area-5. Conv , PreConv , and ProConv represent different local aggregation, PE refers to Positional Encoding, and HPE denotes the proposed High-dimensional Positional Encoding. Δ represents the difference in mIoU between the current row and the first row. PreConv† indicates the use of Conv in the first ABS stage. Please refer to Sec. V and Sec. V-B for the evaluation metrics.
Methods
OA
mAcc
Param.
FLOPs
TP
PointNet [ 12 ]
68.2
63.4
3.5
0.9
6846
PointCNN [ 72 ]
78.5
75.1
0.6
-
-
PointNet++ [ 13 ]
77.9
75.4
1.5
1.7
3019
DGCNN [ 60 ]
78.1
73.6
1.8
4.8
1093
PRANet [ 73 ]
82.1
79.1
2.3
-
-
PointMLP [ 20 ]
85.7
84.4
13.2
31.4
370
TABLE II: 3D object classification on ScanObjectNN [ 25 ] . The best results are highlighted in bold , while the second-best results are underlined . Results marked with an asterisk (*) indicate released values.
Methods
OA
mAcc
Param.
FLOPs
TP
PointNet [ 12 ]
89.2
86.2
3.5
0.9
6848
PointNet++ [ 13 ]
91.9
-
1.5
1.7
3019
PointCNN [ 72 ]
92.2
88.1
0.6
-
-
PointConv [ 14 ]
92.5
-
18.6
-
-
DGCNN [ 60 ]
92.9
90.2
1.8
4.8
1093
DeepGCN [ 74 ]
93.6
90.9
2.2
3.9
629
TABLE III: 3D object classification on ModelNet40 [ 26 ] .
Methods
S3DIS Area-5
S3DIS 6-fold
Param.
FLOPs
TP
mIoU
OA
mAcc
mIoU
OA
mAcc
M
G
ins./sec.
PointNet++ [ 13 ]
53.5
83.0
-
54.5
81.0
67.1
1.0
7.2
332
PointCNN [ 72 ]
57.3
85.9
63.9
65.4
88.1
75.6
0.6
-
-
KPConv [ 47 ]
67.1
-
72.8
70.6
-
79.1
15.0
-
-
RepSurf-U [ 75 ]
68.9
90.2
76.0
74.3
89.9
82.6
1.0
1.0
-
ASSANet [ 23 ]
68.0
89.7
74.3
-
-
-
115.6
36.2
-
TABLE IV: Scene semantic segmentation on S3DIS [ 4 ] .
Fig. 6: Representative visual comparison of semantic segmentation results on S3DIS Area-5. “ST” denotes Stratified Transformer.
Methods
mIoU
Param.
FLOPs
TP
PointNet++ [ 13 ]
53.5
1.0
7.2
332
RepSurf-U [ 75 ]
70.0
1.0
1.0
-
Point Transformer [ 11 ]
70.6
7.8
5.6
-
MinkowskiNet [ 9 ]
72.2
37.9
4.7
-
Fast Point Transformer [ 54 ]
72.5
37.9
-
-
Stratified Transformer [ 24 ]
74.3
18.8
16.1
-
TABLE V: Scene semantic segmentation on ScanNet V2 [ 27 ] val. set.
Methods
mIoU
Param.
FLOPs
SPVNAS [ 79 ]
64.7
10.8
64.5
Cylinder3D [ 80 ]
64.3
55.8
134.2
WaffleIron [ 81 ]
68.0
-
-
SphereFormer [ 78 ]
67.8
32.3
89.8
RangeFormer [ 82 ]
67.6
24.3
-
MinkowskiNet [ 9 ]
63.8
37.9
-
TABLE VI: Scene semantic segmentation on SemanticKITTI [ 28 ] val. set.
Methods
Cls. mIoU
Ins. mIoU
Param.
FLOPs
TP
PointNet [ 12 ]
80.4
83.7
3.6
4.9
-
PointNet++ [ 13 ]
81.9
85.1
1.1
5.5
-
DGCNN [ 60 ]
82.3
85.2
1.3
12.4
-
KPConv [ 47 ]
85.1
86.4
-
-
-
PointMLP [ 20 ]
84.6
86.1
-
-
-
DTNet [ 83 ]
-
86.2
-
-
-
TABLE VII: 3D object part segmentation on ShapeNetPart [ 29 ] .
Fig. 7: Representative qualitative comparison of semantic segmentation on ScanNet. Visualizations for Fast Point Transformer (FPT) exclude points annotated with the ignore label.
TABLE VIII: 3D object detection results on SUN RGB-D validation set with [email protected]. ‡ donates that the model is implemented on MMDetection3D [ 84 ] .
Networks( Ce , B )
HPE
OA
mAcc
mIoU
△
Param.
FLOPs
TP
HPENet_dv (32, [0, 0, 0, 0])
82.7 ± 0.6
58.3 ± 0.7
51.5 ± 0.2
-
0.5
0.8
509
HPENet_dv (32, [1, 1, 1, 1])
85.8 ± 0.5
65.5 ± 0.3
58.8 ± 0.6
+7.3
1.6
1.2
452
HPENet_dv (32, [1, 1, 1, 1])
PE
89.3 ± 0.3
73.4 ± 0.3
66.7 ± 0.4
+15.2
1.6
1.4
328
HPENet_dv (32, [1, 1, 1, 1])
HPE
90.0 ± 0.1
74.5 ± 0.8
68.1 ± 0.4
+16.6
1.8
3.2
309
HPENet_dv ⋆
83.5 ± 0.5
60.1 ± 0.8
53.2 ± 0.7
-5.6
1.6
1.2
507
TABLE IX: Ablation study on S3DIS Area-5 demonstrating the efficacy of ABS-REF view, and contribution of HPE modules. Δ represents the difference in mIoU between the current row and the first row.
Fig. 8: Qualitative comparisons among PointMetaBase [ 5 ] (first row), HPENet (second row), and Ground Truth (third row) on ShapeNetPart [ 29 ] for 3D object part segmentation.
TABLE X: 3D object detection on ScanNet V2 validation set. ‡ donates that the model is implemented on MMDetection3D [ 84 ] .
Line
Type
ABS
REF
BFM
OA
mAcc
mIoU
△
Param.(M)
FLOPs(G)
TP(ins./sec.)
1
87.5 ± 0.4
70.4 ± 0.5
63.9 ± 0.3
-
15.4
10.4
176
2
HPE
√
90.2 ± 0.1
76.1 ± 0.4
70.0 ± 0.4
+6.1
15.8
13.7
160
3
HPE
√
91.0 ± 0.1
76.9 ± 0.6
70.7 ± 0.5
+6.8
15.8
14.5
148
4
HPE
√
√
91.5 ± 0.2
77.9 ± 0.3
72.2 ± 0.2
+8.3
16.0
17.7
137
5
HPE
√
√
√
91.5 ± 0.1
78.4 ± 0.5
72.3 ± 0.2
+8.4
16.1
18.2
128
6
PE
√
√
90.9 ± 0.2
77.0 ± 0.5
70.9 ± 0.5
-1.3
15.4
10.8
144
TABLE XI: Ablation study for positional encoding and backward fusion module with HPENet V2-XL on S3DIS Area-5 justifying HPE use in both ABS and REF stages. Δ in lines 1 through 5 indicates the difference in mIoU between the corresponding row and the first row, while Δ in lines 6 through 8 represents the difference in mIoU between the corresponding row and the fourth row.
Dimension
OA
mAcc
mIoU
△
Param.
FLOPs
TP
3
90.9 ± 0.2
77.0 ± 0.5
70.9 ± 0.5
-
15.3
10.8
144
C/8
91.6 ± 0.2
77.9 ± 0.2
71.9 ± 0.3
+1.0
15.6
14.1
141
C/4
91.5 ± 0.2
77.9 ± 0.3
72.2 ± 0.2
+1.3
16.0
17.7
137
C/2
91.4 ± 0.1
77.4 ± 0.2
71.5 ± 0.1
+0.6
16.7
24.9
129
C
91.4 ± 0.3
78.2 ± 0.7
71.8 ± 0.4
+0.9
18.1
39.3
116
TABLE XII: Ablation study on dimension of HPEMLP with HPENet V2-XL on S3DIS Area-5. ‘C’ denotes the feature channel number. Δ represents the difference in mIoU between the current row and the first row.
Methods
ABS (MLPs)
REF (PT layers)
BFM
OA
mACC
mIoU
△
Param.(M)
FLOPs(G)
Point Transformer [ 11 ]
Local MLPs
PE
90.4
74.8
68.2
-
7.8
5.6
+Non-local MLPs
Non-local MLPs
PE
90.7
76.3
69.7
+1.5
7.8
4.9
+HPE
Non-local MLPs + HPE
HPE
90.5
76.4
70.3
+2.1
7.8
5.2
+BFM
Non-local MLPs + HPE
HPE
√
90.9
76.6
70.7
+2.5
7.9
5.3
Stratified Transformer [ 24 ]
Local MLPs
cRPE
91.7
77.4
71.4
-
8.0
12.3
+Non-local MLPs
Non-local MLPs
cRPE
90.9
78.9
72.0
+0.6
8.0
11.1
TABLE XIII: Ablation of the proposed modules on Point Transformer (PT) [ 11 ] and Stratified Transformer (ST) [ 24 ] for S3DIS Area-5. Δ denotes the mIoU difference relative to the baseline row. cRPE (Contextual Relative Position Encoding) is a spatial relative positional encoding used in ST.
Fig. 9: t-SNE comparison of positional encoding variants for HPENet-dv. Left to right: (a) Baseline, (b) HPE (abs), (c) HPE (mal), (d) HPE.
Fig. 10: Visual comparison of feature-to-label similarity matrices for HPENet V2 variants. The baseline and label matrix are shown for reference.
Fig. 11: Robustness to point density during training and testing on ScanObjectNN across PointNeXt, PointMetaBase, and HPENet V2.
Fig. 12: 3D object classification architecture of HPENet V2. The classification architecture has the same encoder structure as the semantic segmentation model (Fig. 2 in the manuscript).
Fig. 13: More recent representative instantiations of the ABS-REF framework. Abbreviations include SOP: Symmetric OPeration, OP: aggregation OPeration, Two SST blocks: Two Successive Stratified Transformer Blocks.
Methods
Area
OA
mACC
mIoU
Ceiling
Floor
Wall
Beam
Column
Window
Door
Chair
Table
Bookcase
Sofa
Board
Clutter
HPENet V2-S
1
90.8
85.9
77.4
96.3
96.0
84.5
64.9
60.9
81.4
88.0
73.2
83.0
74.5
61.7
71.8
70.3
2
81.9
66.2
51.9
90.4
86.9
80.1
9.0
44.5
17.5
63.4
49.6
43.5
69.3
49.7
30.0
41.1
3
91.4
85.2
76.5
96.3
98.2
84.4
66.8
60.5
48.8
87.7
72.2
81.5
74.0
68.5
82.1
73.4
4
86.9
71.7
60.9
92.0
97.6
82.5
9.0
45.7
36.5
68.2
66.3
76.1
51.6
56.2
52.7
57.9
5
89.4
72.4
65.6
93.6
98.5
83.4
0.0
32.3
52.7
68.0
81.7
88.8
57.9
72.4
66.2
57.4
6
93.5
90.7
83.6
96.7
97.9
88.2
77.7
77.2
86.3
91.2
79.7
88.2
72.4
76.2
79.3
75.8
TABLE XIV: Detailed semantic segmentation results (%) on S3DIS [ 4 ] with 6-fold cross-validation.
Method
mIoU
car
bicycle
motorcycle
truck
other-vehicle
person
bicyclist
motorcyclist
road
parking
sidewalk
other-ground
building
fence
vegetation
trunk
terrain
pole
traffic-sign
HPENet V2-S
66.9
95.5
49.9
79.3
85.7
52.2
75.3
90.8
3.6
93.9
49.9
82.8
0.7
90.0
62.8
90.0
72.7
79.4
65.5
50.4
HPENet V2-B
67.0
95.7
43.8
77.3
91.7
55.0
75.6
91.5
2.1
94.0
49.0
82.7
6.8
90.6
66.0
88.5
70.4
75.4
65.4
51.6
HPENet V2-L
67.4
96.2
52.7
77.1
88.1
64.2
73.6
88.3
0.1
94.0
54.9
82.0
0.4
91.1
67.6
88.6
70.3
75.8
64.8
51.4
HPENet V2-XL
69.3
96.5
55.3
78.8
92.8
66.2
78.4
86.1
25.2
94.6
48.6
83.9
0.3
91.7
69.2
87.7
69.8
74.0
65.4
51.7
TABLE XV: Detailed semantic segmentation results (%) on SemanticKITTI.
Fig. 14: Representative qualitative results of HPENet V2 and the balanced MLP-based method PointMetaBase [ 5 ] on S3DIS Area-5. The main differences are highlighted in red frames.
Fig. 15: Representative qualitative results of HPENet V2 and the balanced MLP-based method PointMetaBase [ 5 ] on S3DIS Area-5. The main differences are highlighted in red frames.
Fig. 16: Representative qualitative results of HPENet V2 and the balanced MLP-based method PointMetaBase [ 5 ] on ScanNet V2. The main differences are highlighted in red frames.
Modern neural architectures for 3D point cloud processing contain both convolutional layers and attention blocks, but the best way to assemble them remains unclear. We analyse the role of different computational blocks in 3D point cloud networks and find an intuitive behaviour: convolution is adequate to extract low-level geometry at high-resolution in early layers, where attention is expensive without bringing any benefits; attention captures high-level semantics and context in low-resolution, deep layers more efficiently, where convolution inflates the parameter count. Guided by this design principle, we propose a new, improved 3D point cloud backbone that employs convolutions in early stages and switches to attention for deeper layers. To avoid the loss of spatial layout information when discarding redundant convolution layers, we introduce a novel, parameter-free 3D positional encoding, PointROPE. The resulting LitePT model has 3.6× fewer parameters, runs 2× faster, and uses 2× less memory than the state-of-the-art Point Transformer V3, but nonetheless matches or outperforms it on a range of tasks and datasets. Code and models are available at: https://github.com/prs-eth/LitePT.
Yuanwen Yue, Damien Robert, Jianyuan Wang +4
ETH Zurich · University of Oxford · University of Zurich
We present SLNet, a lightweight backbone for 3D point cloud recognition designed to achieve strong performance without the computational cost of many recent attention, graph, and deep MLP based models. The model is built on two simple ideas: NAPE (Nonparametric Adaptive Point Embedding), which captures spatial structure using a combination of Gaussian RBF and cosine bases with input adaptive bandwidth and blending, and GMU (Geometric Modulation Unit), a per channel affine modulator that adds only 2D learnable parameters. These components are used within a four stage hierarchical encoder with FPS+kNN grouping, nonparametric normalization, and shared residual MLPs. In experiments, SLNet shows that a very small model can still remain highly competitive across several 3D recognition tasks. On ModelNet40, SLNet-S with 0.14M parameters and 0.31 GFLOPs achieves 93.64% overall accuracy, outperforming PointMLP-elite with 5x fewer parameters, while SLNet-M with 0.55M parameters and 1.22 GFLOPs reaches 93.92%, exceeding PointMLP with 24x fewer parameters. On ScanObjectNN, SLNet-M achieves 84.25% overall accuracy within 1.2 percentage points of PointMLP while using 28x fewer parameters. For large scale scene segmentation, SLNet-T extends the backbone with local Point Transformer attention and reaches 58.2% mIoU on S3DIS Area 5 with only 2.5M parameters, more than 17x fewer than Point Transformer V3. We also introduce NetScore+, which extends NetScore by incorporating latency and peak memory so that efficiency can be evaluated in a more deployment oriented way. Across multiple benchmarks and hardware settings, SLNet delivers a strong overall balance between accuracy and efficiency. Code is available at: https://github.com/m-saeid/SLNet.
Mohammad Saeid, Amir Salarpour, Pedram MohajerAnsari +1
Sirjan University of Technology, Sirjan, Iran · Clemson University, Clemson, SC, USA
Fine-tuning pre-trained point-cloud backbones typically updates all parameters, resulting in substantial computation and memory overhead. More importantly, modern point backbones rely on aggressive tokenization and downsampling, which yields compact global tokens but irreversibly discards fine-grained local geometry, an inherent bottleneck for parameter-efficient adaptation. Consequently, existing PEFT methods that operate only on these coarsened tokens can modulate global semantics but struggle to recover the missing multi-scale locality. We present Point Ladder Tuning (PLT), a locality-aware PEFT framework that performs hierarchical, instance-conditioned adaptation while keeping the backbone frozen. PLT forms a lightweight closed loop: (i) a Hierarchical Ladder Network (HLN) constructs a multi-resolution local feature pyramid directly from raw points; (ii) a Local-Global Fusion (LGF) aligns and fuses local pyramids with intermediate backbone semantics; and (iii) a Dynamic Prompt Generator produces instance-aware multi-scale prompts to modulate the frozen backbone effectively. For dense prediction, we further introduce a lightweight segmentation head that progressively upsamples fused features and leverages backbone priors to refine fine structures. Extensive experiments on classification and dense prediction show that PLT consistently surpasses prior PEFT baselines with minimal tunable parameters. PLT achieves state-of-the-art performance using only 2.71% trainable parameters for classification and 7.69% for dense prediction, and scales favorably to larger backbones, requiring merely 0.36% parameters on PointGPT-L. The code is released at https://github.com/JunLinChang/ECCV2026-PLT.