FLASH: Real-Time LiDAR Super-Resolution via Frequency-Aware Multi-Scale Fusion

June Moh Goo, Zichao Zeng, Jan Boehm

Dept. of Civil, Environmental and Geomatic Engineering, University College London

Overview

FLASH (Frequency-aware LiDAR Adaptive Super-resolution with Hierarchical fusion) is a dual-domain framework for LiDAR range-image super-resolution. Unlike prior spatial-only transformers such as TULIP, FLASH processes both the spatial and the frequency domain, capturing fine-grained geometry and periodic scanning patterns at log-linear cost. It reaches state-of-the-art accuracy on KITTI in a single deterministic pass at real-time speed (66 FPS, 15 ms) — beating even uncertainty-enhanced baselines that need many forward passes.

Highlights

• Dual-domain processing — combines local spatial window attention with global frequency-domain (FFT) analysis.
• Two key modules:
  • Frequency-Aware Window Attention (FA): spatial window attention + a global frequency branch via FFT, fused with a learnable weight.
  • Adaptive Multi-Scale Fusion (MSF): replaces plain skip concatenation with position-specific, multi-scale (1×1/3×3/5×5) feature aggregation, refined by CBAM attention.
• Real-time & single-pass — 66 FPS / 15 ms, vs 7.5 FPS for TULIP with Monte-Carlo Dropout, while being more accurate.
• Implicit uncertainty handling — the frequency branch suppresses noise without stochastic inference.
• State of the art on KITTI across all reconstruction metrics.

Method

FLASH is a Swin-Transformer U-Net for range images (16×1024 → 64×1024, 4× vertical upsampling), built on the TULIP backbone with two additions at every block and every skip connection:

Frequency-Aware Window Attention (FA). Each block runs a spatial window-attention branch in parallel with a frequency branch. The frequency branch takes the 2D FFT, applies a learned spectral gate, and returns via the inverse FFT. The two branches are combined as Output = Attention_spatial + α · FreqBranch(X), with α initialised to 0.1 and learned.

Adaptive Multi-Scale Fusion (MSF). At each skip connection, encoder and decoder features are combined and passed through three parallel convolutions (1×1, 3×3, 5×5); a softmax network predicts position-specific fusion weights, and CBAM refines the result.

Results

KITTI (16×1024 → 64×1024, 4× upsampling). MAE and Chamfer distance (CD) lower is better; voxel IoU / F1 higher is better.

Method	MAE ↓	CD ↓	IoU ↑	F1 ↑
LiDAR-SR	0.7947	0.2992	0.2089	0.3433
SwinIR	0.5776	0.1874	0.3627	0.5300
TULIP	0.4354	0.1342	0.3819	0.5502
FLASH (Ours)	0.3899	0.1161	0.3972	0.5661

Performance by distance range. FLASH is best at both near and far range, with the largest relative gains in the harder far range (30–60 m).

Method	Range	MAE ↓	CD ↓	IoU ↑	F1 ↑
SwinIR	Near (0–30 m)	0.299	0.059	0.394	0.563
LiDAR-SR	Near (0–30 m)	0.443	0.118	0.229	0.371
TULIP	Near (0–30 m)	0.255	0.049	0.418	0.587
FLASH (Ours)	Near (0–30 m)	0.239	0.041	0.434	0.603
SwinIR	Far (30–60 m)	3.064	2.921	0.092	0.168
LiDAR-SR	Far (30–60 m)	4.149	5.347	0.010	0.019
TULIP	Far (30–60 m)	2.302	2.196	0.091	0.167
FLASH (Ours)	Far (30–60 m)	2.045	1.911	0.105	0.189

Efficiency vs. uncertainty modelling. FLASH beats TULIP with Monte-Carlo Dropout (20 samples) in accuracy while running ~9× faster in a single pass.

Method	MC Dropout	MAE ↓	CD ↓	IoU ↑	F1 ↑	Time (ms) ↓
TULIP		0.4354	0.1342	0.3819	0.5502	14
TULIP	✓	0.4070	0.1250	0.3853	0.5538	134
FLASH (Ours)		0.3899	0.1161	0.3972	0.5661	15

Ablation — contribution of each module.

MSF	FA	MAE ↓	CD ↓	IoU ↑	Pre ↑	Re ↑	F1 ↑
		0.4392	0.1348	0.3678	0.5355	0.5356	0.5355
✓		0.3904	0.1209	0.3721	0.5433	0.5368	0.5400
✓	✓	0.3899	0.1161	0.3972	0.5708	0.5617	0.5661

Both modules contribute; adding FA on top of MSF gives the largest IoU jump (0.372 → 0.397). At runtime FLASH reaches 66 FPS.

Qualitative comparison

Three challenging KITTI scenarios: (1) noise suppression around the sensor-mounting region (LiDAR-SR sprays spurious points, TULIP leaves scattered artifacts, FLASH stays clean); (2) edge preservation on large vehicles (baselines blur or smooth the boundaries, FLASH keeps sharp discontinuities); (3) fine-detail recovery of a van's rear-window frame.

Getting Started

Installation

git clone https://github.com/junegoo94/FLASH.git && cd FLASH
conda create -n flash python=3.9 -y && conda activate flash
pip install -r requirements.txt

Data

FLASH trains on paired low-/high-resolution range images (.npy, one range value per pixel, in metres). Arrange each split as flat directories of paired files (same filename in both):

data/kitti/
├── train/{low_res, high_res}/*.npy     # low_res: 16×1024, high_res: 64×1024
└── val/{low_res, high_res}/*.npy

KITTI is 16×1024 → 64×1024 (4× vertical). See the TULIP preprocessing for building range images from raw KITTI.

Pretrained checkpoints

wget https://huggingface.co/jmgoo1118/FLASH/resolve/main/flash_kitti.pth

See CHECKPOINTS.md for details.

Inference

Super-resolve a low-resolution range image and back-project it to a point cloud:

python inference.py \
    --checkpoint flash_kitti.pth \
    --input path/to/low_res.npy \
    --dataset kitti \
    --out_dir out --save_ply

Outputs out/<name>_sr.npy (high-resolution range image, metres) and, with --save_ply, out/<name>_sr.ply.

Training

python train.py \
    --dataset kitti \
    --data_root data/kitti \
    --batch_size 16 --epochs 600 --lr 1.5e-4 \
    --output_dir checkpoints/

train.py trains the full FLASH model (Frequency-Aware attention + Multi-Scale Fusion). The L1 log-range loss is computed inside the model. Checkpoints are written every --save_every epochs.

Citation

If you find this work useful, please cite:

@INPROCEEDINGS{11630360,
  author={Goo, June Moh and Zeng, Zichao and Boehm, Jan},
  booktitle={2026 IEEE International Conference on Image Processing (ICIP)}, 
  title={Flash: Real-Time Lidar Super-Resolution Via Frequency-Aware Multi-Scale Fusion}, 
  year={2026},
  volume={},
  number={},
  pages={1-6},
  keywords={Laser radar;Superresolution;Timing;Printing;Distance measurement;Frequency;Windows;Transformers;Conferences;Learning (artificial intelligence);Lidar super-resolution;Range sensing;Perception;Point Clouds;Deep learning},
  doi={10.1109/ICIP61757.2026.11630360}}

@article{goo2025flash,
  title   = {Real-Time LiDAR Super-Resolution via Frequency-Aware Multi-Scale Fusion},
  author  = {Goo, June Moh and Zeng, Zichao and Boehm, Jan},
  journal = {arXiv preprint arXiv:2511.07377},
  year    = {2025}
}

Acknowledgements

This work was supported by the EPSRC through an industrial CASE studentship with Ordnance Survey (EP/X524840/1, EP/W522077/1). The backbone builds on TULIP (Yang et al., CVPR 2024).