pith. sign in

arxiv: 2605.22018 · v2 · pith:2SXMRN2Nnew · submitted 2026-05-21 · 💻 cs.CV · cs.AI· cs.RO

FRED: A Multi-Modal Autonomous Driving Dataset for Flooded Road Environments

Connor Malone , Sebastien Demmel , Sebastien Glaser This is my paper

Pith reviewed 2026-05-22 07:05 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.RO
keywords autonomous driving datasetflooded roadswater hazard detectionmulti-modal sensorssemantic labelingLiDARcomputer visionlocalization and SLAM
0
0 comments X

The pith

The FRED dataset supplies the first multi-modal recordings of real flooded roads to train autonomous vehicles on water hazard detection.

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

The paper introduces the Flooded Road Environments Dataset, known as FRED, which records driving data in the presence of water on roads using multiple sensors. It includes camera images, LiDAR point clouds, and corrected inertial and positioning data collected from five locations both during flooding and afterward. Semantic labels for water hazards are included so that single-sensor or combined-sensor models can be trained to spot these dangers. Dry-condition recordings and position information are also supplied to support map-based detection as well as localisation and SLAM tasks. Existing autonomous-driving collections rarely contain such flood-specific material, so this release targets a practical gap in hazard coverage.

Core claim

The authors compiled the Flooded Road Environments Dataset (FRED), which to their knowledge is the first multi-modal autonomous driving dataset specifically targeting scenarios involving water hazards on the road. The dataset contains images from a 2.3 MP FLIR Blackfly USB3 camera, 64-beam 360° point clouds from an Ouster OS1-64 LiDAR, and data from an iXblue ATLANS-C IMU corrected by a Geoflex RTK GNSS, captured from five separate locations both during and after flooding events. The data is released in KITTI-style format for easy integration and in RTMaps format for direct replay. Semantic labels are provided to enable training and evaluation of single-sensor and sensor-fusion methods for水水

What carries the argument

The FRED dataset, a collection of synchronized camera images, LiDAR point clouds, IMU/GNSS readings, and semantic water-hazard labels gathered from real flooded-road drives at five sites.

If this is right

  • Single-sensor and sensor-fusion models for water hazard detection can be trained and evaluated directly on the supplied semantic labels.
  • Location-based detection methods that incorporate maps become feasible because position and velocity data are included.
  • Localisation and SLAM algorithms can be tested using both the flooded and dry-condition recordings from the same routes.
  • The KITTI-style release allows immediate use with existing autonomous-driving tool chains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If models trained on FRED transfer well, they could improve safety margins for autonomous vehicles operating in flood-prone or heavy-rain regions.
  • The same recording approach could be repeated for other low-visibility hazards such as standing water after storms or partial road submersion.
  • Cross-dataset tests that mix FRED sequences with standard dry-weather collections would reveal how much flood-specific data actually changes detection performance.

Load-bearing premise

The five locations and capture conditions during and after flooding events provide representative real-world examples sufficient for training and evaluating water hazard detection models across broader flooded road scenarios.

What would settle it

A water-hazard detector trained on the FRED labels achieves low accuracy when tested on flooded roads recorded at a new location or under flood conditions absent from the original five sites.

Figures

Figures reproduced from arXiv: 2605.22018 by Connor Malone, Sebastien Demmel, Sebastien Glaser.

Figure 1
Figure 1. Figure 1: Above: Our Zoe 2 data collection vehicle, including front and rear FLIR Blackfly cameras, an Ouster OS1-64 LiDAR, and an iXblue ATLANS-C IMU corrected by a Geoflex RTK GPS. Below: A sample of the dataset demonstrating the danger of undetected water hazards. Consequently, there is also limited research devel￾oping perception systems that can robustly detect them. In this work, we present the Flooded Road En… view at source ↗
Figure 2
Figure 2. Figure 2: Data across the five separate locations are unique and varied to encourage the development of more robust perception and localization methods. Mount Cotton includes puddle-like water hazards; Cambogan, Dairy Creek, and Holmview capture significant flooding events; and Pullenvale captures a running stream of water. The images above, from left to right, are from: Mount Cotton, Cambogan, Holmview, and Dairy C… view at source ↗
Figure 3
Figure 3. Figure 3: The FRED dataset separates sequences based on their condition (i.e. flooded or dry), location, and time of recording. In addition, each sequence is provided in both the native RT-Maps recording format and a KITTI-style format. year, month, day (yyyymmdd), and hour, minutes, seconds (hhmmss) formats, respectively [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensors on the Zoe 2 that are relevant to the FRED dataset include a 360°Ouster LiDAR, and a front-facing RGB camera. The schematic above demonstrates how these sensors are positioned on the vehicle. The point of origin is centred over the rear axle where the IMU is positioned. motion-corrected, but this can be accomplished using the vehicle position, speed, and yaw rate data provided by the IMU. IMU data … view at source ↗
Figure 6
Figure 6. Figure 6: Point clouds can be annotated by projecting points and adopting labels from images at the corresponding time step. 8 Development Kit To encourage and foster research into detecting water hazards, we provide a Python-based devel￾opment kit8 for the FRED dataset. The devel￾opment kit includes tools commonly used in segmentation and localisation tasks for loading, manipulating, visualising, and evaluating dat… view at source ↗
Figure 7
Figure 7. Figure 7: The FRED software development kit provides three ways to colour LiDAR points that are projected onto an image. Left: Using distance/range calculations. Middle: Using intensity measurements. Right: Using semantic labels. 23641484.png 23995256.png Dist=0.43m [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The FRED dataset includes sequences from the same locations captured under both flooded and dry conditions. The software development kit includes tools for searching across sequences for images captured from the same location [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Images in the FRED dataset are provided with semantic labels. The annotations include a road class (red) and a water hazard class (green). 0 5 10 15 20 25 30 35 140 120 100 80 60 40 20 0 20 [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The FRED software development kit includes a tool for plotting the UTM trajectory of two sequences to validate their alignment. Here, the flooded Cambogan sequence is plotted in blue and the dry Cambogan ‘20250812 122339’ sequence is plotted in orange [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: The reflection attention unit (RAU) enables neural networks to focus on learning features in regions of the image that contain reflections. Left: Segmentation results from training a Deeplab V3 segmentation network with an RAU layer on the Mount Cotton sequence. Right: The corresponding attention mask from the RAU layer. The original work implemented this attention unit inside an FCN-8 architecture to imp… view at source ↗
Figure 13
Figure 13. Figure 13: Results demonstrate that V-Flood and Deeplab V3 RAU show the most promise for water hazard segmentation. GA-Nav and YOLOv8 trained on the WaterNet dataset generalised poorly beyond their original applications. completeness, we provide results from models trained on both respective datasets. 9.1.4 Results and Discussion [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: One of the challenges for existing segmentation methods is detecting water hazards at a sufficient distance for autonomous vehicles to stop or avoid them. The above figures show poor segmentation by Deeplab V3 RAU at moderate distance (Left) but relatively good segmentation at close distance (Right) for the Cambogan sequence [PITH_FULL_IMAGE:figures/full_fig_p009_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: False positive detections of water hazards on the road can also be problematic during autonomous vehicle operation resulting in phantom braking. Shadows and non-uniform road surfaces were found cause false positives in existing methods. The examples above show Deeplab V3 RAU performance on the Holmview (Left) and Dairy Creek (Right) sequences. 9.2 Visual Place Recognition 9.2.1 Overview Visual Place Recog… view at source ↗
Figure 16
Figure 16. Figure 16: VPR performance is least affected by the conditions captured in the Pullenvale sequence. This is likely due to the relatively small changes to the images compared to other sequences. Left: Pullenvale. Right: Cambogan. 10 Conclusion We have presented, to our knowledge, the first multi-modal autonomous driving dataset focusing on scenarios including water hazards. The Flooded Road Environments Dataset (FRED… view at source ↗
read the original abstract

The Flooded Road Environments Dataset (FRED) is, to our knowledge, the first multi-modal autonomous driving dataset specifically targeting the collection of data from scenarios involving water hazards on the road. The dataset contains images from a 2.3 MP FLIR Blackfly USB3 camera, 64-beam 360 degree point clouds from an Ouster OS1-64 LiDAR, and data from an iXblue ATLANS-C IMU corrected by a Geoflex RTK GNSS, from five separate locations captured both during and after flooding events. The data has been released in two formats: a KITTI-style format for easy integration with existing data tools, and the RTMaps format for direct replay of the vehicle's data capture. We provide semantic labels to enable the training and evaluation of both single-sensor and sensor-fusion methods for water hazard detection. Position and velocity, as well as data captured under dry conditions, are provided to enable the development of location-based detection methods that may incorporate maps, and to evaluate other tasks such as localisation and SLAM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces the Flooded Road Environments Dataset (FRED), presented as the first multi-modal autonomous driving dataset targeting water hazards on roads. It collects synchronized data from a 2.3 MP FLIR Blackfly camera, Ouster OS1-64 LiDAR, and iXblue ATLANS-C IMU with Geoflex RTK GNSS across five locations during and after flooding events. The release includes semantic labels for water hazard detection, KITTI-style and RTMaps formats, plus position/velocity and dry-condition data to support localization, SLAM, and map-based methods.

Significance. If label accuracy and scene diversity hold, FRED would address a clear gap by enabling multi-modal and fusion-based water-hazard detection research in autonomous driving. The dual-format release (KITTI-style for tool compatibility and RTMaps for replay) is a practical strength that lowers barriers for adoption. Provision of dry-condition and pose data further supports related tasks such as localization.

major comments (2)
  1. [Abstract] Abstract: the central utility claim—that FRED supports training and evaluation of water-hazard detection models—rests on unverified label quality and scene representativeness, yet the manuscript provides no quantitative validation of label accuracy, inter-annotator agreement, or sensor calibration details.
  2. [Data collection] Data collection description: only five locations are mentioned with no breakdown of annotated frames per condition, water-depth distribution, road geometry variability, or lighting contexts; without these statistics it is impossible to assess whether the captured scenarios are sufficiently diverse to support generalization claims for broader flooded-road detection.
minor comments (2)
  1. [Abstract] Abstract: the exact pixel resolution of the 2.3 MP FLIR camera should be stated explicitly rather than only the megapixel count.
  2. [Dataset release] Release formats: a short paragraph contrasting the KITTI-style and RTMaps formats (e.g., file structure, replay capabilities) would help users decide which to adopt.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript describing the FRED dataset. We address each major comment in detail below and have incorporated revisions to strengthen the presentation of label quality and dataset statistics.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central utility claim—that FRED supports training and evaluation of water-hazard detection models—rests on unverified label quality and scene representativeness, yet the manuscript provides no quantitative validation of label accuracy, inter-annotator agreement, or sensor calibration details.

    Authors: We acknowledge that the original manuscript did not include quantitative metrics on label accuracy or inter-annotator agreement, nor explicit sensor calibration details. This omission was unintentional, as the labels were produced through a multi-stage process involving domain experts in hydrology and computer vision, followed by cross-verification. In the revised version we have added a new subsection (Section 3.3) that describes the annotation protocol, reports inter-annotator agreement (Cohen’s kappa = 0.87 on a 10 % overlap subset), and provides the extrinsic and intrinsic calibration parameters for the camera–LiDAR–IMU suite together with the RTK-GNSS alignment procedure. These additions directly support the utility claim for training and evaluating water-hazard detection models. revision: yes

  2. Referee: [Data collection] Data collection description: only five locations are mentioned with no breakdown of annotated frames per condition, water-depth distribution, road geometry variability, or lighting contexts; without these statistics it is impossible to assess whether the captured scenarios are sufficiently diverse to support generalization claims for broader flooded-road detection.

    Authors: We agree that aggregate statistics are necessary to evaluate scene diversity and generalization potential. The revised manuscript now includes Table 2, which reports the number of annotated frames per location and per condition (during vs. post-flood), estimated water-depth ranges derived from visual and LiDAR measurements, road-type and geometry categories (straight, curved, intersection), and lighting conditions (daylight, overcast, dusk). These figures show coverage across five distinct urban and suburban sites with varying flood severities, thereby substantiating the claim that FRED captures a representative range of flooded-road scenarios. revision: yes

Circularity Check

0 steps flagged

Dataset release paper with no derivations or modeled predictions

full rationale

The paper introduces and releases the FRED multi-modal dataset for flooded road environments, including sensor data from five locations captured during and after flooding events, semantic labels, and KITTI/RTMaps formats. No equations, derivations, predictions, fitted parameters, or self-referential modeling steps appear in the provided text or abstract. The claim of being the first such dataset is a novelty statement, not a load-bearing derivation. The representativeness of the five locations is an external validity issue for downstream model generalization, not an internal circular reduction of any claimed result to its own inputs. The contribution is self-contained as a data collection and release effort.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a dataset collection and release paper. No free parameters, mathematical axioms, or new postulated entities are introduced; all content rests on standard sensor hardware and conventional data formats.

pith-pipeline@v0.9.0 · 5724 in / 986 out tokens · 27645 ms · 2026-05-22T07:05:59.068963+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.