WeatherKITTI: KITTI Weather Benchmarks

• WeatherKITTI is a family of KITTI-derived adverse-weather benchmarks that simulate fog, snow, and rain for robust LiDAR perception evaluation.
• It employs physics-based weather simulations and controlled sensor corruptions to create distinct datasets for tasks like place recognition and denoising.
• Different constructions, such as ResLPR and TripleMixer variants, offer tailored protocols to measure restoration in range images and point-cloud noise suppression.

Searching arXiv for papers on WeatherKITTI and closely related KITTI-derived adverse-weather benchmarks. arxiv_search(query="WeatherKITTI KITTI adverse weather LiDAR benchmark", max_results=10) WeatherKITTI denotes a set of KITTI-derived adverse-weather benchmarks used to study how autonomous-driving perception pipelines behave when originally clear-weather sensor data are corrupted by fog, snow, rain, or related environmental effects. In the recent literature, the name is not attached to a single canonical dataset; rather, it appears in distinct but overlapping benchmark constructions, most prominently a LiDAR place-recognition benchmark built on KITTI odometry and a large-scale LiDAR denoising dataset derived from SemanticKITTI. Related work extends the same KITTI-based adverse-weather paradigm to rainy lane-depth estimation, 3D single-object tracking, depth completion and denoising, and flooded-road perception (Kuang et al., 16 Mar 2025, Zhao et al., 2024).

1. Terminology and scope

A recurrent source of confusion is terminological. One line of work introduces WeatherKITTI as one half of the ResLPR benchmark for LiDAR-based place recognition, whereas another introduces Weather-KITTI as a large-scale synthetic LiDAR denoising dataset with point-wise weather labels. Both inherit KITTI sensor data and impose controlled adverse-weather corruptions, but they differ in representation, annotation granularity, intended downstream task, and evaluation protocol (Kuang et al., 16 Mar 2025, Zhao et al., 2024).

Designation	Primary task	Core contents
WeatherKITTI	LiDAR place recognition	156,951 range images, clean database C00, corrupted query variants S00/F00/R00, loop-closure labels
Weather-KITTI	Point-cloud denoising	22 SemanticKITTI sequences, raw LiDAR matrices $P_i\in\mathbb{R}^{n_i\times 5}$, automatic point-wise weather labels, 130,656 frames
Related KITTI-derived weather resources	Depth, tracking, multi-modal robustness	RainKITTI, KITTI-A, AURORA-KITTI, and broader benchmark proposals including flood data

This multiplicity matters because results reported under the WeatherKITTI name are not directly interchangeable. A place-recognition score such as Recall@1 or $\mathrm{mSR}_l$ and a denoising score such as mIoU measure different failure modes, even when the underlying weather simulators are similar. This suggests that “WeatherKITTI” is best understood as a benchmark family or naming convention rather than a single immutable corpus.

2. Construction principles and weather simulation

In the ResLPR formulation, WeatherKITTI is built from the KITTI odometry dataset’s 64-beam Velodyne HDL-64E scans. Each raw scan is projected into a $64\times1920$ 2D range image with two channels per pixel, encoding the normalized range $d$ and intensity $i$. Training sequences for restoration and robustness tuning are KITTI sequences 03–10, while sequence 00 is used at test time as the clean database and its corrupted variants serve as query sets (Kuang et al., 16 Mar 2025).

The weather corruption process is explicitly physics-inspired. Fog is modeled through backscatter and attenuation using a hard-target versus soft-target formulation,

$\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$

with

$$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$

Snowfall follows a scattering-and-occlusion model,

$$i_{\rm snow} = T_R + i_{\max}\,f_s\, \bigl|\,f_o - (1 - R^*/R_{\max})\bigr|^2, \qquad \hat{\mathbf p}_s = \Bigl(\tfrac{r_s}{\gamma}\,(x,y,z),\,i_{\rm snow}\Bigr),$$

and rain is generated by Monte Carlo particle simulation, with rain-rate $r_r$ controlling probabilistic occlusion and scattering (Kuang et al., 16 Mar 2025).

The TripleMixer Weather-KITTI dataset adopts a related but more point-centric construction. It reuses all 22 SemanticKITTI odometry sequences (00–21), collected with a Velodyne HDL-64E operating at 64 channels, 10 Hz, $\sim120$ m max range, $\mathrm{mSR}_l$0 cm range precision, and vertical resolution $\mathrm{mSR}_l$1. For each sweep, it preserves a matrix

$\mathrm{mSR}_l$2

containing $\mathrm{mSR}_l$3, while also retaining the original timestamp, reflectivity, and full GPS/IMU pose logs. Synthetic weather injection is applied to every frame for light, moderate, and heavy snow, fog, and rain, using explicit parameter ranges: snowfall $\mathrm{mSR}_l$4, $\mathrm{mSR}_l$5, $\mathrm{mSR}_l$6 mm/h; fog $\mathrm{mSR}_l$7, $\mathrm{mSR}_l$8, $\mathrm{mSR}_l$9; rain $64\times1920$0, $64\times1920$1, $64\times1920$2 mm/h (Zhao et al., 2024).

A notable distinction is representational fidelity. ResLPR WeatherKITTI converts the LiDAR stream into range images for restoration and descriptor extraction, whereas TripleMixer Weather-KITTI retains raw point-cloud acquisition information, including range and reflectivity, and automatically generates point-wise weather labels. The former is optimized for loop-closure evaluation; the latter is optimized for denoising and semantic filtering.

3. Data organization, labels, and evaluation protocols

For LiDAR place recognition, WeatherKITTI contains 156,951 LiDAR scans represented as $64\times1920$3 range images across all weather types and severity levels. The test protocol uses sequence 00 clean (C00) as the reference database, while the nine corrupted variants—S00, F00, R00 at severities 1, 2, 3—serve as query scans. Each corrupted scan inherits the original KITTI pose, and loop-closure ground truth is defined by proximity within 5 m and $64\times1920$4. The output label for a query scan is the index set of database scans satisfying that loop-closure threshold (Kuang et al., 16 Mar 2025).

The principal metrics in this setting are Recall@K—specifically R@1 and R@5—together with precision-recall AUC and F1 score, aggregated over R@1 results. ResLPR further introduces the mean stability rate $64\times1920$5 as a relative robustness measure: $64\times1920$6 This metric is designed to express degradation relative to clean-weather performance, rather than reporting only absolute retrieval scores (Kuang et al., 16 Mar 2025).

In the denoising-oriented Weather-KITTI dataset, scale and annotation differ substantially. The corpus contains 130,656 frames, with approximately 14,000 frames per (type, severity) slice, and an average of $64\times1920$7–$64\times1920$8 points per frame. The dataset reports global totals of 139 M snow points, 383 M fog points, and 89 M rain points. Weather noise is labeled at point level using category IDs 110 for snow noise, 111 for fog noise, and 112 for rain noise, while original SemanticKITTI classes remain in 0–29. The per-frame file set includes .bin point clouds, .label files, pose.csv, and metadata.json with LiDAR sensor parameters such as focal offset, focal slope, and pulse width $64\times1920$9 (Zhao et al., 2024).

The denoising evaluation protocol uses train/validation/test splits partitioned per weather modality. Its primary metrics are

$d$0

$d$1

The paper notes that Chamfer Distance or per-point SNR could also be reported in principle, but the main protocol is classification-based (Zhao et al., 2024).

4. Restoration and denoising architectures

Within the ResLPR benchmark, adverse weather is not only a source of evaluation corruption but also the target of an explicit preprocessor. ResLPRNet is a plug-and-play LiDAR range-image restoration network intended to remove weather-induced noise and fill missing returns before descriptor extraction. Its backbone contains WaveTransformer (WAT) Blocks in both encoder and decoder. These blocks first apply a discrete wavelet transform,

$d$2

then reconstruct wavelet features by

$d$3

Feature mixing combines spatial self-attention and grouped-convolution channel mixing, after which a transformer layer fuses $d$4 with the mixed features to capture global context. Decoder stages further include ContextGuide (CTG) Blocks with learnable context embeddings $d$5 and softmax-weighted aggregation,

$d$6

to adapt to weather-type-specific degradation. Skip connections preserve spatial detail, and the restoration objective is

$d$7

Training uses Adam with lr = $d$8 for 120 epochs on $d$9 random crops with horizontal and vertical flips (Kuang et al., 16 Mar 2025).

The TripleMixer line of work addresses the same weather corruption problem at point-cloud level. TripleMixer is composed of a Geometry Mixer Layer, a Frequency Mixer Layer, and a Channel Mixer Layer, respectively designed to capture geometric spatial information, extract multi-scale frequency information, and enhance multi-channel feature information of point clouds. The evaluation includes statistical filters—SOR, ROR, DSOR, DROR—general semantic segmenters adapted to binary classification—SalsaNext, Cylinder3D—and specialized denoisers such as WeatherNet, 4DenoiseNet, 3D-OutDet (Zhao et al., 2024).

These two architectures occupy complementary positions in the pipeline. ResLPRNet is range-image restoration tightly coupled to place recognition, whereas TripleMixer is point-cloud denoising with semantic weather labels. The shared premise is that adverse-weather robustness can be improved by front-end correction rather than by modifying only the downstream descriptor or segmentation network.

5. Empirical performance and task-specific extensions

On the ResLPR WeatherKITTI benchmark, clean-weather place-recognition baselines—Scan Context (SC), CVTNet, and LPSNet, all pretrained on clean KITTI—typically achieve R@1 $i$0 and $i$1–$i$2 on clean KITTI 00. Under moderate weather corruption without preprocessing, however, R@1 often falls below 0.2 and F1 to $i$3. For corrupted scans only, SC records approximately AUC $i$4, F1 $i$5, and R@1 $i$6 on snow, with similarly low values on fog and rain; overall $i$7 is approximately 0.31 for SC, 0.19 for CVTNet, and 0.36 for LPSNet. With ResLPRNet preprocessing, the reported gains are much larger: SC+ResLPR reaches snow AUC $i$8 (+0.66), F1 $i$9, R@1 $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$0, R@5 $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$1, with $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$2; CVTNet+ResLPR reaches snow AUC $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$3 (+0.59), R@1 $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$4, with $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$5; and LPSNet+ResLPR reaches snow AUC $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$6 versus 0.03 corrupted and R@1 $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$7, with $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$8. Comparative preprocessing methods, including WeatherNet and TripleMixer, provide only modest gains, with $\hat{\mathbf p}_f = \begin{cases} \bigl(s\,(x,y,z),\,i_{\rm soft}\bigr), & i_{\rm soft}>i_{\rm hard},\[3pt] \bigl(x,y,z,\,i_{\rm hard}\bigr), & \text{otherwise}, \end{cases}$9 up to $$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$0, below ResLPRNet’s $$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$1 stability (Kuang et al., 16 Mar 2025).

On the denoising side, TripleMixer reports that Weather-KITTI supports state-of-the-art denoising evaluation across snow, fog, and rain, with average mIoU = 96.31% and an improvement over prior methods by at least 2–16% in absolute terms. Because the dataset preserves all original clean scene points and labels only introduced weather returns, the task is explicitly formulated as identifying and suppressing synthetic weather noise while maintaining scene structure (Zhao et al., 2024).

The same adverse-weather KITTI logic has been extended to other tasks. RainKITTI synthesizes rainy lane images from KITTI left-camera frames using the RCFLane pipeline, which sequentially adds a rain layer, global darkening, and depth-dependent fog: $$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$2

$$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$3

The dataset contains 820 original KITTI frames and 820 synthesized rainy frames at $$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$4, split into 715 training and 105 test images. The associated DLKPN restores rainy inputs using two predicted pixel-wise kernels, and its reconstruction score reaches SSIM 0.941 and PSNR 30.601 with 152.9 ms runtime; in depth estimation, KPNDepth(D) reports Abs Rel 0.066, Sq Rel 0.447, RMSE 2.897, RMSE log 0.136, and $$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$5 (Shi, 2024).

For LiDAR tracking, KITTI-A applies synthetic fog, rain, and snow to the KITTI tracking split at five discrete intensity levels per weather, producing 15 variants of each original sequence. It retains Car and Pedestrian categories and reports 1,800 sequences / 96,360 frames for Car and 930 sequences / 91,320 frames for Pedestrian. Evaluation uses Success, Precision, and a Degradation Rate

$$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$6

where $$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$7 (Zhao et al., 13 Jan 2025).

A broader multimodal extension appears in AURORA-KITTI, which introduces 82,177 aligned RGB–LiDAR pairs with metric dense-depth ground truth across {day, night} $$i_{\rm hard}=i\exp(-2\alpha\,R_0), \qquad i_{\rm soft}=i\,R_0^2\,\beta\,i_t.$$8 {clear, fog, rain, snow}, with three severity levels for fog, rain, and snow. It defines Depth Completion and Denoising (DCD) as a unified task and reports that DDCD achieves RMSE 1,799.55 mm, MAE 566.31 mm, iRMSE 4.13, and iMAE 1.93 on AURORA-KITTI zero-shot evaluation (Wang et al., 16 Mar 2026).

The most explicit real-world counterpart to these synthetic weather resources is FRED, a multimodal flooded-road dataset released in both KITTI-style and RTMaps formats. It comprises 10 sequences from five locations, with approximately 5,340 time-synchronized samples, FLIR images, Ouster OS1-64 LiDAR, and IMU/GNSS data. Its authors position it as the “flood” component of a broader WeatherKITTI benchmark, thereby extending the weather-benchmark idea beyond simulated rain, fog, and snow (Malone et al., 21 May 2026).

6. Significance, misconceptions, and limitations

The main scientific role of WeatherKITTI-style resources is controlled robustness evaluation. By parameterizing severity and preserving either loop-closure labels, point-wise weather classes, or dense depth targets, these datasets make it possible to isolate failure modes induced by attenuation, backscatter, scattering, occlusion, and spurious returns. This control is especially clear in the ResLPR benchmark’s use of clean reference databases plus corrupted query sets, and in TripleMixer’s preservation of all original clean scene points alongside exact synthetic-noise labels.

A common misconception is that WeatherKITTI names a single benchmark with a single protocol. The literature instead presents multiple task-specific instantiations: range-image place recognition, point-cloud denoising, lane depth estimation, 3D tracking, and unified depth completion. As a result, benchmark numbers are not directly comparable across papers because the modalities, annotation schemas, and metrics differ. This suggests that precise citation of the variant—ResLPR WeatherKITTI, TripleMixer Weather-KITTI, or a related extension—is methodologically important.

Another limitation concerns the relation between synthetic and real adverse weather. The dominant WeatherKITTI constructions are simulation-based, using physics-inspired or physics-based weather injection on clear KITTI data. Parallel work on real-weather transfer and real-world collection indicates why this distinction matters. In source-only cross-weather LiDAR segmentation, training on SemanticKITTI and evaluating on the real-weather SemanticSTF benchmark yields 24.4 mIoU for a source-only baseline, 36.3 mIoU for data-centric augmentation, 38.9 mIoU for class-centric regularization, and 39.15 mIoU for a geometry-aware regularization method, showing that robustness still depends on transfer beyond synthetic corruption alone (Cheong et al., 3 Nov 2025). Likewise, AURORA-KITTI explicitly concludes that weather-aware, physically consistent data contributes more to robustness than architectural modifications alone, and FRED supplies real flooded-road data rather than simulated water hazards (Wang et al., 16 Mar 2026, Malone et al., 21 May 2026).

Taken together, these developments position WeatherKITTI not merely as a dataset name but as a methodological template: start from a standard KITTI-derived corpus, inject or collect adverse-weather conditions with controlled semantics, and evaluate robustness in a task-specific manner. The continuing proliferation of related benchmarks suggests an ongoing shift from clear-weather accuracy toward explicit all-weather stress testing across localization, recognition, denoising, tracking, segmentation, and depth reconstruction.