Learning to Suppress SPAD-based LiDAR Flare

Linghao Shen; Xuanya Zhu

REVIEW 3 major objections 6 minor 41 references

Learning to Suppress SPAD-based LiDAR Flare

T0 review · 3 major / 6 minor · reviewed 2026-07-12 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SPAD LiDAR flare is suppressed by segmenting first and second echoes plus ambient light as separate modalities with a physics-informed network.

desk verdict Solid multi-echo segmentation for a real SPAD LiDAR failure mode; the 79% mIoU and depth gains are real on their data, but Table 2 only measures mask agreement under the second-echo swap assumption. read the letter →

arxiv 2607.03247 v1 pith:VZ5KETPH submitted 2026-07-03 cs.CV

Xuanya Zhu , Linghao Shen This is my paper · ORCID

classification cs.CV

keywords FlareSuppressionSPADLiDARSemanticSegmentationmulti-echopile-uprangeviewautonomousdriving

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SPAD-based LiDAR is sensitive enough for autonomous driving, but bright retro-reflectors flood neighboring pixels and create flare: false first-echo returns that invent or stretch obstacles in the point cloud. Existing fixes stay at the hardware or rule-based signal level and do not generalize. This paper reframes the problem as semantic segmentation of raw multi-echo measurements. Standard segmenters ignore the fact that a corrupted first echo often still has a usable second echo; the authors therefore build a seven-channel range-view input from both echoes and ambient illumination, then fuse them with depth bins, modality-specific encoding, and adaptive echo weighting. Their model, PILF, reaches 79.32% mean IoU on a new real-world FLARE dataset and, after simple mask-guided echo replacement, cuts depth error in flare regions by roughly 85%. The result is a software-level correction that restores geometric fidelity without changing the sensor.

What carries the argument

PILF (Physically-Informed segmentation for LiDAR Flare): a two-branch range-view network that bins multi-echo signals by depth (Depth Decomposition), extracts modality-specific geometric-photometric features without cross-modal aliasing (Physics-Aware Encoder), and adaptively weights first-echo reliability against complementary second-echo cues under ambient guidance (Echo-Aware Fusion), followed by a directional-attention U-Net decoder.

What would settle it

Collect frames in which both first and second echoes are known to be corrupted or no reliable later echo exists, apply PILF’s mask-guided replacement, and check whether depth error still falls by the reported ~85%; if residual error stays high or valid structure is erased, the central substitution claim fails.

Watch

Extended reading notes

Core claim

Flare suppression for SPAD LiDAR can be cast as three-class semantic segmentation of multi-echo measurements. When first-echo, second-echo, and ambient illumination are treated as distinct modalities and fused by depth decomposition, physics-aware encoding, and echo-aware fusion, the resulting masks reliably mark flare, its bright core, and background. Those masks enable echo-level correction that restores plausible depth while preserving valid structure, reaching 79.32% mIoU on the FLARE dataset and substantially lowering depth error after substitution.

Load-bearing premise

The method assumes that, for most flare-corrupted pixels, a usable second echo still records the true surface and that human labels cleanly separate flare from the bright core that caused it.

Editorial extensions

If this is right

Predicted flare masks can replace corrupted first echoes with second echoes inside the raw stream, cutting depth MSE in flare regions by roughly 86%.
The pipeline runs in the 2D range-view domain at about 37 ms per frame on a single GPU, remaining compatible with 10 fps acquisition.
Because correction is applied at the output level, it can sit alongside existing hardware and signal-level mitigation without altering acquisition.
Feeding the same multi-echo plus ambient representation into ordinary 2D segmenters already improves their flare and core IoU, showing the representation itself carries the physical cue.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

When later echoes are themselves saturated or missing, the same mask would need a residual estimator or third-echo fallback rather than hard substitution.
The multi-echo fusion pattern is likely reusable for other SPAD artifacts such as multipath ghosts once full histograms are unavailable to perception modules.
A safety stack that already consumes range-view features could absorb the flare mask as an extra channel without rebuilding a 3D pipeline.
Synthetic physics renderers of pile-up and crosstalk could stress-test whether annotation boundaries systematically bias the reported mIoU.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This is a clean engineering paper on a concrete automotive SPAD problem. The new pieces are the FLARE dataset (multi-echo range-view with flare/core/background labels), the task framing itself, and the PILF stack (depth-bin decomposition, modality-specific PAE, echo-aware fusion) that actually uses first + second echo + ambient instead of throwing the later returns away.

What they do well: the physics motivation is not hand-waving. Figure 2–3 and the PCA separability claim make the multi-echo case concrete; the 7-channel construction is simple and right for the sensor; ablations on N, ψ, modalities, FlareAug, and the three modules are present and move the numbers in the expected direction; and they beat a reasonable set of 2D/3D/RV baselines under matched training. The qualitative depth maps after mask-guided swap look better, and the latency note (37 ms on a 3090) is honest. Citations cover both the SPAD hardware literature and modern segmentation; no obvious citation games.

Soft spots, in proportion. The stress-test is correct but not fatal: Table 2’s reference is annotation-guided application of the same Eq. 10 swap, so the 84–86% MSE/MAE drops measure how well the predicted mask matches the human flare labels under the assumption that a usable second echo exists. The paper itself flags missing/corrupted later echoes as rare failure cases and supplies no frequency count, no “withhold second echo” ablation, and no independent geometric GT. Single-sensor (IMX459) data and no public code/data links in the manuscript are the other practical limits. None of these break the central empirical claim on the reported test set; they just bound how far you can extrapolate the geometric recovery numbers.

Who it is for: anyone building SPAD or multi-return LiDAR perception who currently discards later echoes. Worth a serious referee; the contribution is real and the experiments are sufficient for a conference track. I would engage, cite the dataset/task if FLARE is released, and treat the depth-recovery percentages with the caveat above.

Referee Report

3 major / 6 minor

Summary. The paper reformulates SPAD-based LiDAR flare suppression as three-class semantic segmentation (flare, core, background) and introduces PILF, a range-view network that treats first-echo, second-echo, and ambient illumination as modalities. It uses Depth Decomposition, a Physics-Aware Encoder, and Echo-Aware Fusion, then applies the predicted mask to replace flare-contaminated first echoes with second echoes (Eq. 10). On the new FLARE dataset (IMX459), PILF reports 79.32% mIoU, outperforming 2D/3D/range-view baselines (Table 1), with ablations on modalities, FlareAug, and modules (Tables 3–4) and large depth MSE/MAE reductions after mask-guided correction (Table 2).

Significance. If the claims hold, this is a useful systems contribution for safety-critical SPAD LiDAR: it moves flare mitigation from configuration-dependent hardware/signal rules to a learnable, multi-echo range-view pipeline that is complementary to sensor-level fixes and runs at ~37 ms/frame. Strengths include a new real-world annotated FLARE dataset, a physically motivated multi-echo input (Fig. 2–4), systematic baselines under native inputs (Table 1), and structured ablations (Tables 3–4). The work is practically oriented and falsifiable on held-out scenes; the main open risk is how well mask quality plus second-echo replacement translate to true geometric recovery when later echoes are missing or corrupted.

major comments (3)

§3.3 Eq. (10) and Table 2: depth recovery is evaluated against “annotation-guided echo correction,” which applies the same first→second echo replacement using the ground-truth flare mask. The reported ~84–86% MSE/MAE reductions therefore measure agreement of the predicted mask with annotators under the assumption that a usable second echo is the correct substitute, not independent geometric accuracy. This overstates recovery relative to real 3D fidelity. Please reframe Table 2 as a mask-consistency metric for the correction rule, and/or add an independent depth reference (e.g., co-registered non-SPAD LiDAR, multi-view stereo, or controlled targets) on a subset of scenes.
§4.2 Practical Discussion and Eq. (10): the pipeline assumes a clean second echo exists for flare pixels. The paper notes that “multiple corrupted echoes or missing reliable later echoes” occur but gives no frequency. Without quantifying the fraction of flare (and core-adjacent) pixels that lack a valid second echo—and without a test-time ablation that withholds or corrupts x2—the suppression claim and Table 2 numbers remain optimistic. Please report this fraction on FLARE and show depth/mask metrics stratified by second-echo availability.
§4.1 Dataset / generalization: all data come from a single IMX459 sensor; train is 90% of main-road scenes (~2k frames) and test mixes the remaining 10% with other scenes (~500 frames). Table 1’s large gains are therefore within one sensor and a partially related scene distribution. For the claim of an “effective solution … across diverse sensing scenarios” (abstract), either add a held-out scene type / different SPAD configuration split with separate metrics, or clearly limit the claim to this sensor and acquisition setup and discuss transfer risk.

minor comments (6)

Table 1 caption/body: “regios” → “regions”; also clarify why RangeFormer is listed with dashes only (no code) rather than omitted.
§4.1 Normalization Eq. (11): λ is “sensor-dependent … fixed,” but no value or sensitivity is given; a one-line note or ablation would help reproducibility.
Fig. 8: qualitative depth maps support the story, but adding failure cases (over-/under-correction, missing second echo) next to successes would balance the Practical Discussion.
Tables 3–4 report means ± std over five runs; Table 1 does not. Adding variance or a simple significance check for the main mIoU gap would strengthen the headline comparison.
§3.2: ψ(·) and directional attention (horizontal only) are motivated by dataset-specific flare anisotropy; briefly state whether this is sensor/layout-specific so readers know when to retune.
Related Work: hardware/signal methods are covered; a short contrast with any commercial SPAD flare heuristics (if public) would help position the learning baseline.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: standard supervised segmentation against independent human labels, with physics used only as architectural inductive bias.

full rationale

The paper's load-bearing claims are empirical: PILF is trained and evaluated as a three-class semantic segmenter on the held-out portion of the newly collected FLARE dataset (human pixel-wise annotations of flare/core/background derived from intensity/depth/pulse-width maps). The reported 79.32% mIoU (Table 1) and ablations (Tables 3-4) are ordinary supervised metrics against those labels; the network does not redefine its own target. Multi-echo ordering, depth binning (Eqs. 2-3), PAE, and EAF are architectural priors that improve feature separability (Fig. 3) but are not used to define the evaluation metric or the ground-truth masks. The post-hoc suppression step (Eq. 10) simply replaces the first echo with the second wherever the predicted mask is positive; Table 2 then measures the resulting depth maps against the identical replacement performed with the annotated mask. That comparison quantifies mask agreement under the second-echo assumption; it is not a first-principles prediction that reduces to a fitted constant or a self-definition. There are no uniqueness theorems, no self-citation chains that force the result, and no parameters fitted on a subset and then re-presented as independent predictions. The derivation chain is therefore self-contained and non-circular.

Assumptions & free parameters 6 free parameters · 4 assumptions · 4 invented entities

The central empirical claim rests on standard supervised-segmentation practice plus a small set of physics-motivated modeling choices and a handful of architectural hyper-parameters chosen by ablation. No new physical constants or unobserved particles are postulated; the free parameters are ordinary network and binning knobs.

free parameters (6)

N (depth bins) = 16
Number of depth bins in Depth Decomposition; ablated and set to 16 for best mIoU.
PAE base channels C = 16
Channel width of the Physics-Aware Encoder; set to 16.
EAF channels = 64
Channel width after fusion, matched to U-Net backbone; set to 64.
learning rate / iterations / batch size = 1e-3 / 2e4 / 16
Adam 1e-3, 2e4 iterations, batch 16; standard training knobs that affect final numbers.
FlareAug crop/fuse probability = 0.75
Each operation applied with 75% probability to combat 100:2:1 class imbalance.
sensor response coefficient λ = hardware-fixed
Used in intensity-to-energy inverse mapping; fixed by hardware configuration.

assumptions (4)

domain assumption When the first echo is dominated by flare, a later (second) echo frequently contains the geometrically valid return.
Stated in §3.1 Multi-Echo Characteristics and used as the justification for both the input representation and the echo-replacement rule (Eq. 10).
domain assumption Flare and its core exhibit spatial continuity within limited depth ranges while noise is randomly distributed.
Motivates Depth Decomposition (§3.2).
domain assumption Human annotators using intensity, depth and pulse-width maps can reliably label flare versus core versus background.
Dataset construction (§4.1); all supervised metrics rest on these labels.
standard math Standard cross-entropy + Dice loss and U-Net + directional attention are adequate optimizers for the three-class range-view task.
Common CV practice; no novel loss derivation.

invented entities (4)

Physics-Aware Encoder (PAE)
purpose: Extract modality-specific geometric-photometric features while suppressing physically unstable responses via a meta-function gate.
Architectural module introduced in §3.2; no independent physical existence outside the network.
Echo-Aware Fusion (EAF)
purpose: Adaptively weight first-echo versus depth-binned second-echo features with spatial and bin-wise gates.
Architectural module introduced in §3.2.
Depth Decomposition (DD)
purpose: Bin multi-echo returns by first-echo depth range to aggregate coherent structure and disperse noise.
Pre-processing module introduced in §3.2.
FLARE dataset
purpose: Provide pixel-wise annotated multi-echo SPAD frames for supervised flare segmentation.
New data resource claimed released; currently the sole evaluation corpus.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning to Suppress SPAD-based LiDAR Flare." pith.science (2026). https://pith.science/paper/VZ5KETPH

@misc{pith2026260703247,
  author       = {Pith},
  title        = {Pith review of: Learning to Suppress SPAD-based LiDAR Flare},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VZ5KETPH}},
  note         = {Machine review of arXiv:2607.03247}
}

read the original abstract

Single-Photon Avalanche Diode (SPAD)-based Light Detection and Ranging (LiDAR) is emerging for autonomous vehicles due to its high sensitivity and precise depth sensing capabilities. However, flare caused by excessive photon returns or pile-up effects can lead to incorrect depth estimation and exaggerated boundaries in point clouds, resulting in severe distortions of geometric measurements, making flare suppression essential for safety-critical applications. Existing flare mitigation methods primarily operate at the hardware or signal-processing levels. While effective under specific configurations, they are largely rule-based and configuration-dependent, lacking learnable representations that generalize across diverse sensing scenarios. In this work, we reformulate flare suppression as a semantic segmentation problem, enabling data-driven learning of geometric and photometric cues directly from SPAD measurements. We first benchmark representative segmentation models on the newly introduced SPAD flare dataset and observe that they struggle to exploit the intrinsic multi-echo characteristics of SPAD signals. Motivated by this observation, we propose Physically-Informed segmentation for LiDAR Flare (PILF), a learning-based approach that treats the first and second echoes, together with ambient illumination, as distinct modalities, aggregating cross-echo information while jointly encoding geometric and photometric features. Experiments across multiple real-world scenes demonstrate that PILF significantly outperforms compared segmentation models, achieving up to 79.32% mIoU, and providing an effective solution for SPAD-based LiDAR flare suppression.

Figures

Figures reproduced from arXiv: 2607.03247 by the authors.

**Figure 1.** Illustration of SPAD LiDAR flare effects in point clouds. (a) The captured 3D point cloud. (b) The point cloud after flare suppression. (c) The depth map. (d) The ambient illumination image. Excessive photon returns from retro-reflective surfaces, such as traffic signs, can produce strong flare that leads to false or exaggerated structures in the point cloud, potentially distorting geometric measurements. This high… view at source ↗

**Figure 2.** Illustration of SPAD flare formation. Excessive photons cause flare to occupy the first echo in neighboring pixels, while valid returns appear in later echoes [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 4.** Comparison of conventional and proposed input construction pipelines. Traditional processing retains only the first echo. Our method selects two echoes and incorporates ambient illumination to construct a 7-channel representation. counts, depth corresponds to time of flight, and pulse width measures the temporal spread of the returned pulse. We further approximate flare as an excessive illumination component and i… view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: Overview of PILF. (a) The overall architecture processes first echo x_1 , second echo x_2 , and ambient illumination x_0 through two branches. (b) Depth Decomposition (DD) organizes multi-echo signals into depth-aware bins. (c) Physics-Aware Encoder (PAE) extracts moda…

**Figure 6.** Figure 6: Example of depth decomposition. Within a specific depth range, objects occluded in the second echo, as well as potential flare regions, appear more completely after being merged with the first echo. not strictly aligned in 3D: intensity may integrate multiple reflecti…

**Figure 7.** Figure 7: A flare-core region is cropped from the source image and fused into another image while preserving the correspondence between flare and its associated core. both flare and valid core regions, the dataset supports evaluation of flare suppression in terms of both semant…

**Figure 8.** Figure 8: Qualitative visualization of depth maps after flare suppression. The upper block compares different 2D segmentation methods, while the lower block shows additional scenarios under diverse illumination conditions and flare patterns. PILF achieves more complete flare sup…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 1 linked inside Pith

[1]

In: Learning to Suppress SPAD-based LiDAR Flare 15 Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: To- wards vision transformers for 3d semantic segmentation in autonomous driving. In: Learning to Suppress SPAD-based LiDAR Flare 15 Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 5240–5250 (2023)

2023
[2]

IEEE Access (2024)

Byun, Y.S., Jeong, R.G.: High-speed outlier removal filter for lidar sensor point cloud data. IEEE Access (2024)

2024
[3]

IEEE Sensors Journal21(5), 5656–5667 (2020)

Chen, G., Wiede, C., Kokozinski, R.: Data processing approaches on spad-based d-tof lidar systems: A review. IEEE Sensors Journal21(5), 5656–5667 (2020)

2020
[4]

In: Proceedings of the European conference on computer vision (ECCV)

Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801–818 (2018)

2018
[5]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)

2022
[6]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolu- tional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075–3084 (2019)

2019
[7]

Contributors, M.: Mmsegmentation: Openmmlab semantic segmentation toolbox and benchmark (2020)

2020
[8]

In: Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5–7, 2020, Proceedings, Part II 15

Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5–7, 2020, Proceedings, Part II 15. pp. 207–222. Springer (2020)

2020

Show all 41 references

[9]

Frontiers in Physics10, 906671 (2022)

Cusini, I., Berretta, D., Conca, E., Incoronato, A., Madonini, F., Maurina, A.A., Nonne, C., Riccardo, S., Villa, F.: Historical perspectives, state of art and research trends of spad arrays and their applications (part ii: Spad arrays). Frontiers in Physics10, 906671 (2022)

2022
[10]

In: 2016 46th European Solid-State Device Research Conference (ESSDERC)

Ficorella, A., Pancheri, L., Dalla Betta, G.F., Brogi, P., Collazuol, G., Marrocchesi, P.S., Morsani, F., Ratti, L., Savoy-Navarro, A.: Crosstalk mapping in cmos spad ar- rays. In: 2016 46th European Solid-State Device Research Conference (ESSDERC). pp. 101–104. IEEE (2016)

2016
[11]

In: Pro- ceedings of the IEEE/CVF international conference on computer vision

Gupta, A., Ingle, A., Gupta, M.: Asynchronous single-photon 3d imaging. In: Pro- ceedings of the IEEE/CVF international conference on computer vision. pp. 7909– 7918 (2019)

2019
[12]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition

Gupta, A., Ingle, A., Velten, A., Gupta, M.: Photon-flooded single-photon 3d cam- eras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition. pp. 6770–6779 (2019)

2019
[13]

IEEE Transactions on Electron Devices69(6), 2794–2805 (2021)

Gyongy, I., Dutton, N.A., Henderson, R.K.: Direct time-of-flight single-photon imaging. IEEE Transactions on Electron Devices69(6), 2794–2805 (2021)

2021
[14]

In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11108–11117 (2020)

2020
[15]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Ikeda, K., Hara, R., Nagata, R., Sako, O., Ding, Z., Kado, T., Fujioka, I., Beppu, T., Isogawa, M., Yoshioka, K.: Ghost-fwl: A large-scale full-waveform lidar dataset for ghost detection and removal. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

2026
[16]

Sensors21(13), 4481 (2021)

Incoronato, A., Locatelli, M., Zappa, F.: Statistical modelling of spads for time-of- flight lidar. Sensors21(13), 4481 (2021)

2021
[17]

Optics express26(16), 20622–20632 (2018) 16 X.Zhu et al

Jahromi, S., Kostamovaara, J.: Timing and probability of crosstalk in a dense cmos spad array in pulsed tof applications. Optics express26(16), 20622–20632 (2018) 16 X.Zhu et al

2018
[18]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228–240 (2023)

2023
[19]

Journal of the Optical society of America61(1), 1–11 (1971)

Land, E.H., McCann, J.J.: Lightness and retinex theory. Journal of the Optical society of America61(1), 1–11 (1971)

1971
[20]

IEEE Signal Pro- cessing Magazine37(4), 50–61 (2020)

Li, Y., Ibanez-Guzman, J.: Lidar for autonomous driving: The principles, chal- lenges, and trends for automotive lidar and perception systems. IEEE Signal Pro- cessing Magazine37(4), 50–61 (2020)

2020
[21]

Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer:Hierarchicalvisiontransformerusingshiftedwindows.In:Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021)

2021
[22]

Advances in neural information processing systems32(2019)

Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. Advances in neural information processing systems32(2019)

2019
[23]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431–3440 (2015)

2015
[24]

In: 2019 IEEE/RSJ international conference on in- telligent robots and systems (IROS)

Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on in- telligent robots and systems (IROS). pp. 4213–4220. IEEE (2019)

2019
[25]

In: 2016 fourth international confer- ence on 3D vision (3DV)

Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international confer- ence on 3D vision (3DV). pp. 565–571. Ieee (2016)

2016
[26]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Po, R., Pediredla, A., Gkioulekas, I.: Adaptive gating for single-photon 3d imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16354–16363 (2022)

2022
[27]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652–660 (2017)

2017
[28]

Advances in neural information processing systems30(2017)

Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learn- ing on point sets in a metric space. Advances in neural information processing systems30(2017)

2017
[29]

Optics express16(12), 8381–8394 (2008)

Rech, I., Ingargiola, A., Spinelli, R., Labanca, I., Marangoni, S., Ghioni, M., Cova, S.:Opticalcrosstalkinsinglephotonavalanchediodearrays:anewcompletemodel. Optics express16(12), 8381–8394 (2008)

2008
[30]

In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....

2015
[31]

In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision

Scheuble, D., Holzhüter, H., Peters, S., Bijelic, M., Heide, F.: Lidar waveforms are worth 40x128x33 words. In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision. pp. 28913–28924 (2025)

2025
[32]

JSAP Review 2023, 230402 (2023)

Tashiro, Y., Ito, K.: Spad depth sensor for automotive lidar systems. JSAP Review 2023, 230402 (2023)

2023
[33]

Sensors20(18), 5203 (2020)

Tontini, A., Gasparini, L., Perenzoni, M.: Numerical model of spad-based direct time-of-flight flash lidar cmos image sensors. Sensors20(18), 5203 (2020)

2020
[34]

Advances in neural information pro- cessing systems30(2017) Learning to Suppress SPAD-based LiDAR Flare 17

Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017) Learning to Suppress SPAD-based LiDAR Flare 17

2017
[35]

IEEE Journal of Selected Topics in Quantum Electronics20(6), 299–305 (2014)

Veerappan, C., Charbon, E.: A substrate isolated cmos spad enabling wide spectral response and low electrical crosstalk. IEEE Journal of Selected Topics in Quantum Electronics20(6), 299–305 (2014)

2014
[36]

Sensors21(11), 3839 (2021)

Villa, F., Severini, F., Madonini, F., Zappa, F.: Spads and sipms arrays for long- range high-speed light detection and ranging (lidar). Sensors21(11), 3839 (2021)

2021
[37]

In: CVPR (2024)

Wu, X., Jiang, L., Wang, P.S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., Zhao, H.: Point transformer v3: Simpler, faster, stronger. In: CVPR (2024)

2024
[38]

Advances in neural information processing systems34, 12077–12090 (2021)

Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems34, 12077–12090 (2021)

2021
[39]

Yuan, Y., Chen, X., Wang, J.: Object-contextual representations for semantic seg- mentation.In:ComputerVision–ECCV2020:16thEuropeanConference,Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16. pp. 173–190. Springer (2020)

2020
[40]

Advances in Neural Information Processing Systems34, 10326–10338 (2021)

Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmenta- tion. Advances in Neural Information Processing Systems34, 10326–10338 (2021)

2021
[41]

arXiv preprint arXiv:2008.01550 (2020)

Zhou, H., Zhu, X., Song, X., Ma, Y., Wang, Z., Li, H., Lin, D.: Cylinder3d: An ef- fective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550 (2020)

2008 arXiv

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.

[1] [1]

In: Learning to Suppress SPAD-based LiDAR Flare 15 Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: To- wards vision transformers for 3d semantic segmentation in autonomous driving. In: Learning to Suppress SPAD-based LiDAR Flare 15 Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 5240–5250 (2023)

2023

[2] [2]

IEEE Access (2024)

Byun, Y.S., Jeong, R.G.: High-speed outlier removal filter for lidar sensor point cloud data. IEEE Access (2024)

2024

[3] [3]

IEEE Sensors Journal21(5), 5656–5667 (2020)

Chen, G., Wiede, C., Kokozinski, R.: Data processing approaches on spad-based d-tof lidar systems: A review. IEEE Sensors Journal21(5), 5656–5667 (2020)

2020

[4] [4]

In: Proceedings of the European conference on computer vision (ECCV)

Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801–818 (2018)

2018

[5] [5]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)

2022

[6] [6]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolu- tional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075–3084 (2019)

2019

[7] [7]

Contributors, M.: Mmsegmentation: Openmmlab semantic segmentation toolbox and benchmark (2020)

2020

[8] [8]

In: Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5–7, 2020, Proceedings, Part II 15

Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5–7, 2020, Proceedings, Part II 15. pp. 207–222. Springer (2020)

2020