Relational Feature Distillation for Lightweight 3D Point Cloud Segmentation
Model Overview
TopoPT is a family of lightweight 3D point cloud segmentation models obtained by compressing
LitePT-S through a compact student architecture (
TrimPT) trained with
Stage-wise Relational Feature Distillation (SRFD).
TrimPT reduces LitePT-S's channel widths from (36, 72, 144, 252, 504) to (36, 54, 108, 180, 360) and cuts stage-3 attention depth from 6 to 4 blocks, preserving the full 1024-token attention window. This yields 5.84 M parameters and 12.95 GFLOPs — 2.18× fewer parameters and 1.96× fewer FLOPs than LitePT-S.
SRFD is a training-only objective applied at the compressed attention stages (stages 3–4). It matches pairwise cosine-similarity matrices between teacher (frozen LitePT-S) and student (TrimPT) features, explicitly regularizing their local affinity structure. Stage-specific linear projectors align teacher and student channel dimensions before computing both a pointwise cosine loss (L_pw) and a relational Frobenius loss (L_rel). After training, the teacher and all projectors are discarded — TopoPT has the same deployed architecture, checkpoint size, FLOPs, and latency as TrimPT.
On ScanNet semantic segmentation, TopoPT achieves 76.6% mIoU at 5.84 M parameters — matching the official LitePT-S result (76.5%) with substantially fewer resources.
Available Checkpoints
All checkpoints follow the naming convention {dataset}-{task}-{model}-{epochs}.
lw-c = TrimPT (no distillation) · lw-c-kd = TopoPT (with SRFD)
Semantic Segmentation
Instance Segmentation
Ablation Checkpoints (ScanNet, 100 epochs)
These correspond to the compression ablation study (Table 2 in the paper).
Inference Efficiency
Profiled on NVIDIA RTX 3090 (batch size 1, 300 forward passes). Since the teacher and projectors are discarded after training, TopoPT and TrimPT have identical inference cost.
| Dataset | Model | Params (M) | GFLOPs | Latency (ms) | FPS | Mem (GB) | Size (MB) |
|---|
| ScanNet | LitePT-S | 12.71 | 25.42 | 34.08 | 29.34 | 1.332 | 145.8 |
| ScanNet | TrimPT / TopoPT | 5.84 | 12.95 | 30.78 | 32.49 | 1.211 | 67.1 |
| NuScenes | LitePT-S | 12.71 | 25.42 | 35.81 | 27.93 | 0.717 | 145.7 |
| NuScenes | TrimPT / TopoPT | 5.84 | 12.95 | 28.84 | 34.68 | 0.432 | 67.0 |
How to Use
Please refer to the
GitHub repository for full setup, training, and evaluation instructions. The codebase follows the same interface as
LitePT and
Pointcept.
Quick start — loading a TopoPT checkpoint for inference:
1# Install and set up environment following the GitHub README, then:
2export PYTHONPATH=./
3python tools/test.py \
4 --config-file configs/scannet/semseg-lw-c-kd-100epoch.py \
5 --num-gpus 4 \
6 --options save_path=exp/topopt_scannet \
7 weight=/path/to/model_best.pth
Architecture Details
| Property | LitePT-S (teacher) | TrimPT / TopoPT (student) |
|---|
| Channels | (36, 72, 144, 252, 504) | (36, 54, 108, 180, 360) |
| Stage depths | (2, 2, 2, 6, 2) | (2, 2, 2, 4, 2) |
| Attention stages | 3, 4 | 3, 4 |
| Attention window | 1024 tokens | 1024 tokens |
| Parameters | 12.71 M | 5.84 M |
| GFLOPs | 25.42 | 12.95 |
Stages 1–2 use sparse convolution; stages 3–4 use windowed multi-head self-attention with PointROPE positional encoding (same as LitePT). SRFD distillation is applied at stages 3–4 during training only.
Citation
1@inproceedings{topopt2026,
2 title={{Relational Feature Distillation for Lightweight 3D Point Cloud Segmentation}},
3 author={Anonymous},
4 booktitle={...},
5 year={2026}
6}
1@inproceedings{yuelitept2026,
2 title={{LitePT: Lighter Yet Stronger Point Transformer}},
3 author={Yue, Yuanwen and Robert, Damien and Wang, Jianyuan and Hong, Sunghwan and Wegner, Jan Dirk and Rupprecht, Christian and Schindler, Konrad},
4 booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
5 year={2026}
6}
