---

title: >- [Paper Note] Learning to Suppress SPAD-based LiDAR Flare description: >- [ECCV 2026][Autonomous Driving][SPAD LiDAR] Reformulates SPAD LiDAR flare suppression as semantic segmentation and proposes PILF to fuse multi-echo and ambient cues, achieving 79.32% mIoU on FLARE and reducing flare depth MSE by 85.89%. tags: - ECCV 2026 - Autonomous Driving - SPAD LiDAR - Semantic Segmentation - Flare Suppression date: 2026-09-19 content_hash: 278e20876c5094d6

Learning to Suppress SPAD-based LiDAR Flare

Conference: ECCV 2026
Paper: ECCV 2026
Area: Autonomous Driving
Keywords: SPAD LiDAR, Flare Suppression, Multi-Echo Representation, Semantic Segmentation, Physical Priors

TL;DR

Addressing the severe point cloud distortions and spurious obstacles caused by pile-up and optical crosstalk in SPAD LiDARs facing retro-reflective surfaces, this paper reformulates flare suppression as a semantic segmentation task, introduces a physically-informed framework (PILF) that integrates dual echoes with ambient illumination, and restores undistorted scene geometry via localized echo replacement.

Background & Motivation

Single-Photon Avalanche Diode (SPAD)-based LiDAR has emerged as a promising sensing modality for autonomous driving and mobile robotics, offering single-photon sensitivity and precise direct time-of-flight (d-ToF) range measurement. However, when confronting highly retro-reflective objects commonly encountered on roads—such as traffic signs, roadside delineators, and vehicle rear reflectors—the avalanche detectors are overwhelmed by massive photon returns within extremely narrow time windows. This surge triggers severe detector saturation and pile-up effects, accompanied by optical crosstalk spreading across neighboring array pixels. These physical phenomena manifest as extensive "LiDAR flare" artifacts in 3D point clouds, exaggerating object boundaries and fabricating floating obstacles in mid-air, which severely endangers downstream planning and control modules in autonomous driving.

Existing mitigation techniques predominantly operate at the hardware level (e.g., optical isolation trenches, on-chip time gating) or through low-level signal processing (e.g., histogram peak suppression, temporal gating filters). While partially effective in controlled bench tests, these non-learnable methods rely heavily on fixed empirical rules and narrow sensor calibration parameters, exhibiting poor adaptability across diverse atmospheric conditions, vehicle vibration profiles, and dynamic lighting variations. Crucially, conventional LiDAR processing pipelines discard all echoes beyond the first echo to satisfy strict bandwidth and latency budgets. In strong flare regimes, however, the first echo is corrupted by crosstalk photons arriving prematurely, whereas the valid surface reflection of the target object remains preserved within the second echo.

To overcome the brittle nature of handcrafted filters and unlock the full potential of raw SPAD hardware, this work reformulates SPAD LiDAR flare suppression as a data-driven semantic segmentation task over native range-view projections. Core idea: construct a physically grounded 7-channel range-view representation combining first and second echoes with ambient illumination, decouple heterogeneous physical cues via depth decomposition and physics-aware encoding, and restore undistorted geometric point clouds through pixel-level echo replacement guided by predicted flare masks.

Method

Overall Architecture

PILF takes as input a 7-channel cylindrical range-view image natively aligned with the sensor scanning grid: first-echo physical attributes (intensity \(I_1\), depth \(D_1\), pulse width \(P_1\)), second-echo attributes (\(I_2, D_2, P_2\)), and an ambient illumination channel \(L\). Following non-linear energy transformation and min-max normalization, the input is processed along two parallel physical streams: a single-echo stream preserving fine-grained geometric details of the primary return, and a multi-echo stream that applies Depth Decomposition (DD) to discretize depth space into range bins, aggregating spatially coherent structures while dispersing random speckle noise. A shared Physics-Aware Encoder (PAE) extracts decoupled geometric-photometric representations for each modality. An Echo-Aware Fusion (EAF) module then adaptively combines first-echo, depth-binned, and ambient features using learned depth-wise weights and spatial gates. A lightweight U-Net backbone equipped with a horizontally constrained multi-head attention decoder predicts a three-class mask (flare, core, background). Finally, the predicted mask guides echo-level correction on raw sensor data by replacing flare-corrupted first echoes with their corresponding second echoes, outputting clean point clouds.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["7-Channel SPAD Input<br/>(I1,D1,P1, I2,D2,P2, L)"] --> B["Depth Decomposition (DD)<br/>Discretize range into depth-aware bins"]
    A --> C["Single-Echo Direct Stream (x1)"]
    B --> D["Physics-Aware Encoder (PAE)<br/>DWConv + Meta-modulation + GroupConv"]
    C --> D
    D --> E["Echo-Aware Fusion (EAF)<br/>Ambient embedding + Depth weights + Spatial gate"]
    E --> F["Directional Attention U-Net<br/>Horizontally constrained multi-head decoder"]
    F --> G["Echo-Level Mask Replacement<br/>Swap flare first echo with second echo"]
    G --> H["Flare-Suppressed Point Cloud & Depth Map"]

Key Designs

1. Depth Decomposition (DD): Discretizing range space to aggregate coherent physical structures and disperse speckle noise

Multi-echo SPAD signals in autonomous driving environments contain significant random speckle noise along the depth axis that lacks spatial continuity, while valid object surfaces and physical flare artifacts exhibit strong spatial correlation within localized depth intervals. Direct concatenation of raw multi-echo tensors introduces severe noise into the network, destabilizing feature learning. Authors note that retro-reflective cores and their surrounding flares are confined to specific depth bands, and occluded surfaces reappear in subsequent echoes at consistent depth coordinates. PILF divides the dynamic depth range \(D\) of the first echo into \(N\) uniform depth bins (\(N=16\) in experiments).

For each bin \(n\) covering depth range \(d_n = [\frac{(n-1)D}{N}, \frac{nD}{N})\), binary masks \(B^{(n)}\) are computed from first-echo depth \(d(i, j)\) and applied to both first (\(x_1\)) and second (\(x_2\)) echoes: $$ \hat{x}^{(n)} = B_1^{(n)} \odot x_1 + B_2^{(n)} \odot x_2, \quad n=1, \dots, N $$ By aggregating multi-echo measurements within corresponding depth slices, target structures occluded in the first echo are reconstructed, while randomly distributed depth outliers are dispersed across orthogonal bins, boosting signal-to-noise ratio and feature separability.

2. Physics-Aware Encoder (PAE): Decoupling heterogeneous photon responses and mitigating non-linear distortions

On the range-view grid, intensity (photon count), depth (time of flight), and pulse width (temporal pulse spread) share identical pixel coordinates but characterize disparate physical properties along the laser beam. Saturated reflections cause detector pulse broadening, non-linear photon response, and timing jitter, inducing local geometric inconsistencies across modalities. Sharing convolutional kernels directly across all physical channels causes physical aliasing and entangles modality-specific representations.

PAE first processes each physical measurement independently using depth-wise convolutions (DWConv), yielding modality-wise features \(x_{\text{dw}}\). To suppress extreme pulse responses and saturated outliers, an adaptive meta-modulation map \(x_{\text{meta}}\) is generated via a meta-function \(\psi\): $$ x_{\text{meta}} = \psi(\text{Conv}{1 \times 1}(x)) $$ Selecting a Gaussian decay kernel }\(\psi(x) = \exp(-x^2)\) provides sharp yet smooth suppression against extreme saturation values. Modulated features are subsequently integrated via group convolution across channels and combined with original inputs via residual gating: $$ \text{PAE}(x) = x_{\text{dw}} + \text{GroupConv}{3 \times 3}(x $$ This design extracts expressive spatial semantics while preserving the physical integrity of individual sensor channels.}}) \odot x_{\text{meta}

3. Echo-Aware Fusion (EAF): Dynamically balancing multi-echo reliability with ambient illumination guidance

In non-flare regions, the primary echo provides clean, high-fidelity geometry, rendering second-echo inputs redundant; conversely, within flare zones, primary measurements are completely corrupted, and subsequent echoes become the only source of truth. A static fusion scheme fails to handle this dramatic spatial and depth-wise confidence shift. EAF embeds ambient illumination \(x_0\) into contextual guidance \(x_{\text{amb}}\), concatenates it with echo features, and stabilizes sample-level statistics across wide dynamic ranges using convolution and LayerNorm.

Global channel descriptors are then extracted via adaptive max pooling from the refined first echo \(x_{\text{f}}\) and depth-binned features \(\hat{x}_{\text{f}}^{(n)}\): $$ z_{\text{f}} = \text{MaxPool}(x_{\text{f}}), \quad z_{\text{f}}^{(n)} = \text{MaxPool}(\hat{x}{\text{f}}^{(n)}) $$ Along the depth dimension, a learnable linear layer \(\ell\) computes bin-wise relevance coefficients \(w_{\text{d}}^{(n)}\) to emphasize valid depth ranges. Along spatial coordinates, a spatial gate \(w_{\text{e}} = \sigma(\text{Conv}_{3 \times 3}(x_{\text{f}}))\) modulates local confidence based on primary geometry. The final fused representation integrates both mechanisms: $$ x \right) $$ This formulation mirrors photon integration along the depth axis prior to range-view projection, providing optimal feature weighting for downstream segmentation.}} = x_{\text{f}} + w_{\text{e}} \odot \sum_{n=1}^N \left( w_{\text{d}}^{(n)} \hat{x}_{\text{f}}^{(n)

4. Lossless Echo-Level Mask Correction: Native sensor stream restoration via arrival-order physical priors

Once the binary flare mask \(M\) is predicted, point cloud correction is performed directly within the parsed native sensor data stream rather than through computationally intensive 3D spatial filtering or meshing. Because flare originates from optical crosstalk triggering premature avalanche events, echo temporal ordering is inverted in affected pixels. For all pixels flagged as flare (\(M(i, j) = 1\)), PILF replaces the corrupted primary echo with the valid secondary echo: $$ E'{i,j,0,:} = \begin{cases} E $$ Because header metadata, timestamps, and calibration descriptors remain intact, the corrected sensor stream integrates seamlessly into existing vehicle software stacks (e.g., 3D object detection, LiDAR odometry) with zero pipeline disruption.}, & M(i,j)=1 \ E_{i,j,0,:}, & M(i,j)=0 \end{cases

Key Experimental Results

Main Results

Evaluation is conducted on the FLARE benchmark collected with a Sony IMX459 SPAD LiDAR, encompassing urban roads, tunnels, highways, retro-reflective traffic signs, and vehicle lights. Baselines comprise prominent 2D semantic segmentation networks, point-based/voxel-based 3D networks, and native Range View (RV) architectures. Metrics include Flare IoU, Core IoU, and overall mIoU.

Table 1: Quantitative segmentation performance on the FLARE dataset (from original Table 1)

Architecture Type	Method	Input Modalities	Flare IoU (%)	Core IoU (%)	mIoU (%)
2D Segmentation	FCN [Long et al., 2015]	\((I_{1,2}, D_{1,2}, P_{1,2}, L)\)	57.08	69.22	63.15
2D Segmentation	DeepLabV3+ [Chen et al., 2018]	\((I_1, D_1, P_1)\)	63.21	75.34	69.28
2D Segmentation	DeepLabV3+ [Chen et al., 2018]	\((I_{1,2}, D_{1,2}, P_{1,2}, L)\)	67.72	76.56	72.14
2D Segmentation	OCRNet [Yuan et al., 2020]	\((I_1, D_1, P_1)\)	63.85	72.89	68.37
2D Segmentation	OCRNet [Yuan et al., 2020]	\((I_{1,2}, D_{1,2}, P_{1,2}, L)\)	65.69	73.48	69.59
2D Segmentation	SegFormer [Xie et al., 2021]	\((I_{1,2}, D_{1,2}, P_{1,2}, L)\)	52.26	69.92	61.09
2D Segmentation	Mask2Former [Cheng et al., 2022]	\((I_{1,2}, D_{1,2}, P_{1,2}, L)\)	52.95	69.21	61.08
3D Segmentation	PVCNN [Liu et al., 2019]	\((x, y, z, I_1)\)	35.40	42.10	38.75
3D Segmentation	RandLA-Net [Hu et al., 2020]	\((x, y, z, I_1)\)	65.40	68.20	66.80
3D Segmentation	Cylinder3D [Zhou et al., 2020]	\((x, y, z, I_1)\)	70.17	72.57	71.37
3D Segmentation	PTv3 [Wu et al., 2024]	\((x, y, z, I_1)\)	57.01	54.17	55.59
RV Segmentation	RangeNet++ [Milioto et al., 2019]	\((D_1, x, y, z, I_1)\)	48.83	46.69	47.76
RV Segmentation	SalsaNext [Cortinhal et al., 2020]	\((D_1, x, y, z, I_1)\)	65.66	76.74	71.20
RV Segmentation	RangeViT [Ando et al., 2023]	\((D_1, x, y, z, I_1)\)	56.34	57.38	56.86
Ours	PILF (Ours)	\((I_{1,2}, D_{1,2}, P_{1,2}, L)\)	78.58	80.06	79.32

To measure direct geometric restoration, Table 2 compares normalized depth errors before and after PILF flare suppression against human-annotated echo correction.

Table 2: Depth recovery accuracy before and after flare suppression (from original Table 2)

Evaluation Region	Stage	MSE	MAE	Relative Reduction (%)
Flare Region	Before Suppression	0.2275	0.3849	-
Flare Region	After Suppression	0.0321	0.0613	85.89% (MSE) / 84.08% (MAE)
Foreground (Flare + Core)	Before Suppression	0.1447	0.2447	-
Foreground (Flare + Core)	After Suppression	0.0219	0.0413	84.85% (MSE) / 83.11% (MAE)

Ablation Study

Ablations investigate depth bin granularity \(N\), meta-function choices \(\psi(x)\), input modalities, and network modules.

Table 3: Ablation on depth bins \(N\) and meta-function \(\psi(x)\) (from original Table 3)

Depth Bins \(N\)	mIoU (%)	Meta-Function \(\psi(x)\)	mIoU (%)
\(N = 4\)	78.21 ± 0.73	\(1 / (1 + x^2)\)	77.62 ± 0.30
\(N = 8\)	77.87 ± 0.46	\(1 / \exp(x^2)\)	78.49 ± 0.29
\(N = 16\)	78.49 ± 0.29	\(\cos(\frac{\pi}{2} \cdot \tanh(x))\)	77.67 ± 0.41
\(N = 32\)	77.28 ± 0.60	\(1 / (1 + \|x\|)\)	77.83 ± 0.22

Table 4: Stepwise ablation on input modalities and network components (from original Table 4)

Category	Ablation Variable	Experimental Setting	mIoU (%)
Modalities	Primary Echo Only	No \(x_2\), No \(x_0\)	75.18 ± 0.45
Modalities	Dual Echoes	With \(x_2\), No \(x_0\)	75.56 ± 0.81
Modalities	Primary + Ambient	No \(x_2\), With \(x_0\)	77.67 ± 0.22
Modalities	Full Modalities	With \(x_2\), With \(x_0\)	78.49 ± 0.29
Architecture	Vanilla Backbone	Standard Conv replacing PAE/EAF, \(N=1\)	64.41 ± 0.35
Architecture	Add PAE	Backbone + PAE	71.01 ± 0.49
Architecture	Add PAE + EAF	Backbone + PAE + EAF (\(N=16\))	75.40 ± 0.54
Architecture	Full PILF	PAE + EAF + Directional Attention Decoder	78.49 ± 0.29

Key Findings

• Failure of standard 3D perception on hardware artifacts: 3D geometric networks (PVCNN, PTv3) treat corrupted first-echo points as legitimate spatial structures, lacking access to multi-echo timing distributions; consequently, their mIoU lags PILF by 23% to 40%.
• Impact of physical inductive bias: Stepwise integration of PAE (+6.60%), EAF (+4.39%), and horizontally constrained attention (+3.09%) raises performance from 64.41% to 78.49%, proving that decoupled physical encoding and adaptive echo weighting are critical for distinguishing flare.
• Ambient illumination as a key discriminator: Ambient light alone yields a +2.49% mIoU improvement over pure single-echo input (77.67% vs. 75.18%), as it allows the model to differentiate intrinsic surface reflectivity from optical oversaturation.

Highlights & Insights

• Reformulating signal processing as data-driven 2D segmentation: Instead of designing heuristic hardware filters or running heavy 3D point cloud deep nets, PILF reframes the problem into range-view segmentation. Operating at 37 ms/frame on an RTX 3090, it seamlessly achieves 10 fps real-time processing on automotive hardware.
• Physical closed loop via echo replacement: Recognizing that flare is not arbitrary noise but a corrupted temporal pre-trigger, PILF uses semantic masks to swap the bad first echo with the valid second echo, recovering ground-truth scene depth without geometric approximation.
• FlareAug for severe class imbalance: In automotive scenes, background-to-flare-to-core pixel ratios exceed 100:2:1. The proposed crop-and-fuse strategy, customized to the horizontal propagation behavior of SPAD arrays, provides balanced and diverse training supervision.

Limitations & Future Work

• Severe multi-echo corruption: When retro-reflective surfaces are extremely close or intense, optical crosstalk can saturate both primary and secondary echoes. In such cases, echo replacement lacks a clean fallback return, leaving residual depth artifacts.
• Sensor-specific horizontal crosstalk assumption: The directional 1D attention and FlareAug partitioning assume horizontal optical crosstalk typical of the IMX459 architecture. Adapting to sensors with 2D isotropic or diagonal crosstalk requires reconfiguring spatial priors.
• Future directions: Integrating PILF directly into end-to-end multi-modal 3D perception networks (e.g., 3D occupancy and object detection), where predicted flare probabilities serve as confidence weights for point cloud featurization.

Related Work & Insights

• vs. Hardware/Signal-level Mitigation (Chen et al. [3], Jahromi & Kostamovaara [17]): Handcrafted signal filters rely on rigid thresholds that fail under unpredictable environmental lighting; PILF leverages learning-based representations to handle dynamic real-world driving scenes.
• vs. Full-Waveform Denoising (Ghost-FWL [Ikeda et al., 2026], Scheuble et al. [31]): Waveform approaches require transferring full d-ToF histograms over high-bandwidth buses; PILF requires only discrete first/second echoes and ambient data, fitting practical automotive bandwidth constraints.
• vs. 3D LiDAR Segmentation (Cylinder3D [41], SalsaNext [8]): Spatial 3D segmentation ignores the time-of-flight physics of point generation; PILF exploits multi-echo photon arrival ordering in range view to resolve ambiguity with minimal computation.

Rating

• Novelty: ⭐⭐⭐⭐☆ (Pioneering reformulation of SPAD LiDAR flare suppression as multi-modal semantic segmentation with physical echo replacement)
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ (Introduces the benchmark FLARE dataset with extensive 2D/3D baselines, rigorous metric evaluation, and depth restoration verification)
• Writing Quality: ⭐⭐⭐⭐⭐ (Clear physical exposition, comprehensive mathematical formulation, and well-designed figures)
• Value: ⭐⭐⭐⭐⭐ (Directly addresses a critical safety failure mode in automotive SPAD LiDAR with an efficient, production-compatible solution)

Related Papers

• [ICCV 2025] LightsOut: Diffusion-based Outpainting for Enhanced Lens Flare Removal
• [CVPR 2026] Probabilistic Discrepancy Learning for Roadside LiDAR Scene Completion
• [CVPR 2026] LEADER: Learning Reliable Local-to-Global Correspondences for LiDAR Relocalization
• [ICLR 2026] Adaptive Augmentation-Aware Latent Learning for Robust LiDAR Semantic Segmentation
• [CVPR 2026] Learning to Identify Out-of-Distribution Objects for 3D LiDAR Anomaly Segmentation