Zenseact Open Dataset: A large-scale and diverse multimodal dataset for autonomous driving
1. Paper overview
In one sentence: ZOD uses diverse, calibrated driving keyframes and separate temporal recordings to broaden perception research, while its baselines expose unresolved long-range and rare-class failures. e02e03e04e06e08e11e13e14
| At a glance | What to know |
|---|---|
| Research problem | Reader analysis The authors argue that densely sampled driving sequences yield highly correlated examples while leaving geography, weather, and lighting underrepresented. ZOD prioritizes diverse European scenes and high-resolution forward perception, then supplies separate temporal subsets. The resulting design trades dense temporal supervision and surround-camera coverage for broader sampling and detailed keyframe labels. Generalization and safer driving are motivations, rather than demonstrated deployment outcomes. e02e03e08e15 |
| Core mechanism | |
| A key reported result | 3D object detection by distance: mAP/CDS: 0–150 m 0.25/0.46; 0–50 m 0.33/0.62; 50–100 m 0.19/0.33; 100–150 m 0.06/0.08; 150–250 m 0.01/0.02. mAP and CDS, reported decimal scores. CenterPoint on ZOD Frames; range bins in metres, with metrics following the cited Argoverse 2 protocol. Split details and uncertainty are not reported. Within-dataset range breakdown. A footnote reports CenterPoint mAP 0.18 on Argoverse 2 at 0–150 m, compared with ZOD’s 0.25. Performance collapses at long range despite distant labels. The cross-dataset footnote does not isolate dataset quality, because the evaluated data distributions differ. e11e15 |
| Reading caution | Source description Keyframe-only labels do not support consistent object tracking. Full-sequence labels, supporting camera frames for Frames, additional recordings, and high-definition maps are future work. Distant 3D annotation quality and recall diminish; some traffic-sign and ego-road labels are missing. e15 |
Core contributions
Figure 2. The acquisition layout supplies complementary views whose coordinates must be reconciled using calibration. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Locate Cam near the windshield, VLS128 near the roof center, the two VLP16 sensors at the sides, and OxTS nearer the rear axle. Red, blue, and green mark each sensor’s X, Y, and Z axes; crossed and dotted circles distinguish downward and upward directions. These are local sensor frames: the camera and LiDAR arrows do not all point in the same directions as the vehicle reference. Section 3.2 separately specifies the rear-axle reference as X-forward, Y-left, Z-up. The figure and that text therefore describe different coordinate frames, whose relationship is provided by the per-datapoint calibrations. e04e05
What it supports. The dataset’s geometric usefulness depends on calibrated relationships among camera, LiDAR, and navigation measurements. Table 2 identifies an 8 MP front camera and three LiDAR sensors; the diagram shows their placement. This is the acquisition mechanism behind the multimodal annotations, rather than an architecture for a learned controller.
Where the evidence stops. The drawing gives orientation conventions and approximate placement, not numerical transformation matrices or calibration-error estimates. Use the supplied intrinsic/extrinsic calibrations for alignment. Treating all displayed axes as one shared frame would misread the figure.
2. Motivation
2.1 The problem and the proposed response
The authors argue that densely sampled driving sequences yield highly correlated examples while leaving geography, weather, and lighting underrepresented. ZOD prioritizes diverse European scenes and high-resolution forward perception, then supplies separate temporal subsets. The resulting design trades dense temporal supervision and surround-camera coverage for broader sampling and detailed keyframe labels. Generalization and safer driving are motivations, rather than demonstrated deployment outcomes. e02e03e08e15
2.2 What this reading follows
A driving dataset can contain many images while repeatedly showing almost the same scene. Zenseact Open Dataset addresses that problem by separating broadly sampled keyframes from continuous recordings. Its front camera, multiple LiDARs, and navigation system support detailed spatial annotations and recorded ego-motion across European driving conditions. Read this paper as a design for collecting and organizing evidence, followed by tests of how difficult that evidence is to learn from. Geographic coverage is substantial, but the detection and traffic-sign baselines reveal weaknesses hidden by aggregate scores. The anonymization experiment asks a narrower question: whether changing training images alters detection on original images. e02e03e04e06e08e11e13e14
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Datasets |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The dataset classification is supported by the sensor records, spatial annotations, temporal subsets, ego trajectories, and logged controls. Motion-prediction/planning relevance describes possible uses of these data, not demonstrated planning performance. Architecture, prediction paradigm, and quadrant are appropriately not applicable: this paper introduces a data resource and separate perception baselines, rather than an integrated world-action model. e06e08e09e11e13e14
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
5. Method in detail
5.1 Build the sample before choosing the learning problem
Start with a Frame as a single annotated camera keyframe. Its nearby LiDAR scans and longer navigation record add context, but they do not turn it into a densely labeled video clip. Apply the per-datapoint calibration to relate sensor coordinates, then select only the measurements appropriate to the task. For example, navigation after the keyframe can supply an ego-trajectory target. Reader interpretation: a causal prediction experiment must keep that future information out of its observation inputs, even though the data package makes it conveniently available. If the task requires continuous camera observations, move to Sequences; if it requires longer localization trajectories, consider Drives. In both cases, check supervision separately from recording duration. The paper explicitly leaves temporally consistent object annotations to future work. e05e06e08e15
Figure 3. One keyframe links image-space labels, three-dimensional boxes, and recorded ego-motion. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across modalities, beginning with the full camera image at upper left. Boxes indicate objects, while colored road and lane overlays provide dense spatial labels. The lower-left panel magnifies a distant image region to show how small annotated objects become. The right panel displays the corresponding LiDAR scene, 3D boxes, and the 200 m future ego trajectory identified by the caption. Section 3.1 describes navigation measurements as a source of trajectory ground truth. The illustrated path therefore belongs to recorded data; it is not evidence that a learned planner predicted or executed that path. The labels are attached to a keyframe. e06e08e09e15
What it supports. The example makes the dataset’s multi-task structure tangible: one scene connects object perception, road/lane understanding, geometry, and ego-motion. Those aligned signals could support learning across tasks. The visual itself establishes what kinds of data are represented, rather than proving a benefit from a particular multi-task training objective.
Where the evidence stops. Some Frames lack traffic-sign or ego-road labels, and annotation quality declines at long range. The example does not establish complete coverage across the dataset. Surrounding sensor measurements also do not supply temporally consistent object tracks.
5.2 Separate the presence of distant targets from learning them reliably
The acquisition hardware and annotated street scene explain why ZOD can represent small, distant objects. They do not show that an algorithm can recognize those objects reliably. Table 4 supplies the next link: a CenterPoint baseline performs much worse in distant bins. Table 5 adds another dimension, showing that rare traffic-sign classes remain difficult despite strong aggregate classification performance. Reader interpretation: range and class frequency should be examined jointly before attributing failure to either one alone. Figure 7 motivates that caution by showing large class-wise detection disparities. Annotation completeness is a third factor because the authors acknowledge declining 3D label quality and recall at distance. A useful follow-up must therefore control the evaluated class and audit labels, rather than treating the range curve as a clean sensor-resolution ablation. e04e06e11e12e13e15
5.3 Read anonymization as a controlled training-domain comparison
The anonymization experiment changes the images used to train a detector while keeping evaluation on original images. That design asks whether a model trained with transformed faces and license plates transfers to the original evaluation domain. Read each detector’s three rows as a matched comparison of image modes; the reported repeated runs describe variation within each condition. The authors conclude that neither anonymization approach significantly degrades their 2D detection baseline. Reader interpretation: this supports a narrow practical claim, not universal invariance to anonymization. It does not establish what happens when evaluation images are also transformed, when facial detail is itself a target, or when temporal consistency matters. A public-data follow-up using only blurred and DNAT images would be informative, but its evaluation domain would differ from the paper’s original-image protocol. e07e14e15
5.4 Training and inference
During training
Baseline learning comprises CenterPoint for 3D detection, a ResNet-50-based traffic-sign classifier, and Faster-RCNN with FPN/ResNet-50 in Detectron2 plus YOLOv7 for image detection. The anonymization study trains each detector separately on original, DNAT, and blurred images. No shared world/action model, new loss, or frozen-module schedule is specified. e11e13e14
The supplied paper names models and evaluation protocols but does not specify train/validation counts, split construction, optimizer settings, training duration, image resizing, CenterPoint voxel configuration, hardware, or software versions. Exact replication cannot be inferred from the named architectures. e11e13e14
During inference
The reported inference is offline perception: 3D detection evaluated by range and class, traffic-sign classification, and 2D detection evaluated on original images. Logged steering and pedal signals are data, not actions generated or executed by a proposed controller. No planning rollout, feedback controller, or closed-loop driving result is reported. e06e11e12e13e14
5.5 Implementation flow
- Collect complementary measurements
An identical layout across collection vehicles combines one 3848 × 2168 front camera with a 120° horizontal field of view, one VLS128 and two VLP16 LiDARs, and an OxTS RT3000 navigation system. Camera and LiDAR data are recorded at 10 Hz; high-precision GNSS/IMU at 100 Hz. Average combined point clouds contain about 254,000 points. e04e06
- Align and package sensor records
Per-datapoint intrinsic and extrinsic calibration supports transformations between sensor frames. The ISO-8855 reference origin is tied to the chassis at the rear-axle center under calibration load, with X forward, Y left, and Z up. LiDAR files contain timestamp, position, intensity, and diode index; navigation files provide positions, orientation, velocities, and related fields. e04e05e06
- Separate spatial diversity from temporal context
Each Frame has one image in two anonymized versions, LiDAR over ±1 second, and navigation data from five seconds before until 25 seconds after the keyframe or 300 m ahead, whichever occurs first. Sequences contain continuous sensor and vehicle data but annotate only their middle image. Drives extend recording duration for localization and mapping. Scene metadata supports filtering by location, weather, road type, and lighting. e07e08
- Annotate visible structure
Human annotators create labels followed by quality checks. Road segmentation has 15 top-level classes; lane instances also carry properties. Objects visible in both camera and LiDAR receive 3D cuboids with center, dimensions, and quaternion rotation, while image objects receive 2D boxes. The paper calls these 9-DOF boxes; it lists four quaternion parameters. Some Frames lack traffic-sign or ego-road labels. e09e15
6. Experiments & results
Zenseact Open Dataset (ZOD) combines geographically dispersed driving keyframes with shorter and longer continuous recordings. Its contribution is calibrated multimodal data and perception annotations, with baseline experiments exposing weaknesses at long distances and on rare classes. The reviewed release supports research on perception, temporal reasoning, and ego-motion, but its sparse temporal labels and incomplete distant annotations limit what those experiments establish about tracking or autonomous control.
6.1 Read the original evidence
Table 1. ZOD separates broad keyframe coverage from continuous recording, with several comparison quantities requiring their footnotes. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the three ZOD rows: they represent different data products, so their frame counts and durations should not be read as a single continuously labeled video collection. Compare Frames with Waymo Open along geographic coverage and annotated-frame count, then examine camera coverage and the map column to see the tradeoffs. Keep the footnotes in view. The 55.6 hours for Frames counts two-second LiDAR snippets; only the center camera image is supplied. The annotation-range column is a 99.9th-percentile statistic, and the asterisk identifies sequential annotation in other datasets. Section 4.1 defines the geographic metric using a common 75 m radius. e03e10
What it supports. At the common radius, ZOD Frames covers 705 km² versus Waymo Open’s 76 km², with 100,000 versus 400,000 annotated frames. This supports wider geographic sampling per labeled frame in the reported comparison. ZOD’s camera covers 120°, and the table lists no map for its subsets.
Where the evidence stops. The table measures dataset properties, not generalization or driving safety. PandaSet lists 300 m annotation range, so ZOD’s long-range claim must retain its comparable-size qualification. The alternative 150 m-radius calculation on page 8 should not be mixed with the common-radius comparison.
Table 4. Long-distance annotations expose a substantial baseline detection deficit. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the first column as an aggregate over 0–150 m. It overlaps the next three columns and must not be treated as an additional disjoint distance band. The remaining columns separate 0–50, 50–100, 100–150, and 150–250 m. Follow either metric horizontally to see how the same baseline performs farther from the ego vehicle. Section 4.2 names CenterPoint and says both metrics follow the cited Argoverse 2 protocol; the source does not reproduce their full mathematical definitions. Values here are decimal scores. The paper reports no repeated-run uncertainty for this table, unlike its separate anonymization experiment. e11e15
What it supports. CenterPoint mAP drops from 0.33 at 0–50 m to 0.06 at 100–150 m and 0.01 at 150–250 m; CDS falls from 0.62 to 0.08 and 0.02. The dataset contains distant targets that this baseline handles poorly. Wider annotation range is not equivalent to effective detector range.
Where the evidence stops. Section 6 also reports declining label quality and recall at distance. The table does not separate detector limitations from label incompleteness or changing class composition. Its cross-dataset footnote compares different datasets, so it cannot isolate a causal benefit of ZOD.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Geographical coverage of annotated Frames Union of radius-75 m regions around annotated ego-poses; descriptive dataset comparison, not a learned-model test split. | ZOD Frames: 705 km². Diversity area, square kilometres | Waymo Open: 76 km²; Argoverse 2: 17 km²; nuScenes: 5 km². The paper takes the Waymo and nuScenes values from its reference [25]. Supports roughly ninefold coverage over the largest listed comparator under the common radius. The separately reported 2039 km² uses a 150 m radius; its stated 26-fold comparison to the old baseline changes the metric radius. e03e10 |
| 3D object detection by distance CenterPoint on ZOD Frames; range bins in metres, with metrics following the cited Argoverse 2 protocol. Split details and uncertainty are not reported. | mAP/CDS: 0–150 m 0.25/0.46; 0–50 m 0.33/0.62; 50–100 m 0.19/0.33; 100–150 m 0.06/0.08; 150–250 m 0.01/0.02. mAP and CDS, reported decimal scores | Within-dataset range breakdown. A footnote reports CenterPoint mAP 0.18 on Argoverse 2 at 0–150 m, compared with ZOD’s 0.25. Performance collapses at long range despite distant labels. The cross-dataset footnote does not isolate dataset quality, because the evaluated data distributions differ. e11e15 |
| Traffic-sign classification under class imbalance ResNet-50-based classifier on ZOD; classes with fewer than ten signs in the validation set are excluded. Split size and uncertainty are unspecified. | F1-macro 78.5%; F1-micro 95.4%; most-common ten-class accuracy 93.4%; least-common ten-class accuracy 65.4%. F1-macro, F1-micro, and average accuracy for the ten most/least common classes, percent | The class-balanced score and rare-class accuracy are substantially lower than their frequency-dominated/common-class counterparts; no competing classifier is tabulated. Aggregate accuracy conceals a long-tail weakness even after excluding the least-supported validation classes. e13 |
| Training-image anonymization and 2D detection ZOD Frames; train on original, DNAT, or blurred images, evaluate on original images for vehicles, vulnerable vehicles, and pedestrians. Three runs per condition. | Faster-RCNN: original 30.23 ± 0.09, DNAT 30.28 ± 0.03, blur 30.17 ± 0.06. YOLOv7: original 33.62 ± 0.04, DNAT 33.74 ± 0.09, blur 33.67 ± 0.01. COCO AP; mean ± standard deviation across three runs | Compare anonymization modes within each detector against its original-image training baseline. The authors report no statistically significant degradation. The small differences support similar AP in this setting; the paper supplies no named significance test or equivalence margin and does not test other downstream tasks. e14e15 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 5. Rare-class performance is substantially weaker than the aggregate traffic-sign score. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the first two entries together. The paper uses F1-macro as the class-balanced view and F1-micro to summarize aggregate performance; their disagreement signals that a single overall score is inadequate. Then read the two accuracy columns: the upward-ten marker denotes the ten most common classes, while the downward-ten marker denotes the ten least common classes. The arrows identify frequency groups, not directions in which the metric should move. Section 4.3 specifies a ResNet-50-based classifier and excludes classes with fewer than ten validation signs. The table therefore describes the retained validation classes, not every class in the release’s traffic-sign taxonomy. e13
What it supports. F1-micro is 95.4%, whereas F1-macro is 78.5%. Accuracy is 93.4% for the ten most common classes and 65.4% for the ten least common classes. Together, these measurements support a long-tail weakness that a high aggregate score can obscure, even after the validation-support exclusion.
Where the evidence stops. This is a diagnostic comparison of metrics and class groups, not an intervention proving why rare classes fail. The paper does not give uncertainty, retained-class counts, or split sizes here; its exclusion prevents extrapolation to the rarest unsupported classes.
Table 3. Changing the training-image anonymization method produces similar detection AP on original evaluation images. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Within each detector block, compare the original, DNAT, and blur rows before comparing architectures. Mode changes the training images; the caption fixes evaluation to original images in every row. Begin with overall AP, then inspect the additional columns for metric, size, and class breakdowns. The caption identifies the last three class columns as vehicles, vulnerable vehicles, and pedestrians. Every entry reports a mean and standard deviation over three runs. Section 4.4 specifies Faster-RCNN with an FPN and ResNet-50 backbone in Detectron2, alongside YOLOv7. Because the evaluation domain stays fixed, the table isolates a particular training-data comparison rather than all possible anonymization effects. e07e14e15
What it supports. Overall AP is 30.23 ± 0.09, 30.28 ± 0.03, and 30.17 ± 0.06 for Faster-RCNN trained on original, DNAT, and blurred images. YOLOv7 reports 33.62 ± 0.04, 33.74 ± 0.09, and 33.67 ± 0.01. These values support the authors’ observation of little overall change in the tested setting.
Where the evidence stops. The authors describe no statistically significant degradation, but do not name a test or equivalence margin. Standard deviations are not confidence intervals. Effects on other tasks remain untested, and original-image evaluation access is not established by the described public anonymized release.
7. Analysis & limitations
7.1 What the evidence leaves open
Keyframe-only labels do not support consistent object tracking. Full-sequence labels, supporting camera frames for Frames, additional recordings, and high-definition maps are future work. Distant 3D annotation quality and recall diminish; some traffic-sign and ego-road labels are missing. e15
The text generally says annotations reach 245 m, Table 1 defines its range statistic as the 99.9th percentile, and the limitations section uses 250 m. These are not interchangeable guarantees of complete annotation. PandaSet’s 300 m entry also prevents interpreting ZOD’s range claim as an absolute record across every dataset size. e03e11e15
Geographic breadth does not imply balanced conditions: clear weather accounts for 61.2% and daytime for 77.3% of Frames. No geographic holdout experiment establishes generalization. Anonymization evidence concerns 2D detection only; temporal inference, gaze-sensitive tasks, and control remain untested here. e10e14e15
7.2 Questions for discussion
- How much of the long-range detection deficit remains after auditing missing labels and holding object class constant?
- Would a geographically separated evaluation split support the generalization motivation more directly than diversity area?
- How does anonymization affect tasks that depend on facial detail or temporal consistency?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires the paper-era dataset version, per-frame calibrations, annotation taxonomies and missing-label handling, and matching model/evaluation configurations. The main text defers exact taxonomies, compression evidence, and expanded anonymization results to supplementary material. e05e09e16
Exact Table 3 replication requires original evaluation images, while the described public Frame images are DNAT and blurred versions. Original-image access is not established here. A study using anonymized evaluation images must disclose that changed protocol. e07e14
Reader-proposed checks: stratify CenterPoint’s range results by class and audited label completeness, then compare DNAT/blur training with paired seeds on a common evaluation set. The illustrated edition specifies controls and falsifiable outcomes; neither check has been run. e11e12e14e15
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Does the range deficit persist with class and annotation coverage controlled?
Reader-proposed check, not performed: pin a ZOD version, split, CenterPoint configuration, and evaluation protocol, then train once and retain fixed predictions. Evaluate the four disjoint Table 4 range bins separately for each sufficiently represented class. On the same selected frames, compare scores under released labels and independently audited labels, keeping matching rules unchanged. Report audited missing-label counts, class support, mAP, and CDS. If the within-class distant deficit persists after the audit, it supports a detection limitation beyond label completeness. If the pattern changes substantially, annotation coverage contributed materially to the reported range diagnosis. The paper does not provide all configuration details needed for exact replication. e09e11e12e15
Check 2: Test anonymization equivalence under an explicit evaluation domain
Reader-proposed check, not performed: use identical frame splits, labels, initialization seeds, resizing, augmentations, and training budgets for DNAT and blur conditions of one named detector. Evaluate each trained model on both image variants of the same held-out Frames. Predeclare an equivalence margin, for example ±0.5 overall AP points, and estimate paired uncertainty across seeds; also report the three class APs. Equivalence requires the interval to lie within the margin, not merely a nonsignificant difference. Add original-image training/evaluation only if that material becomes available through an authorized source. Without it, label this a public-variant extension rather than an exact reproduction of Table 3. e07e14e15
8.3 Reading coverage
Visual audit: Visually inspected the title/author block and CVF version notice; all pages of the supplied PDF; Figures 1–7; Tables 1–5 including relevant captions, legends, and footnotes; and every final crop. Figure 2 local-axis arrows and upward/downward markers were checked against Section 3.2: the sensor frames differ from the shared vehicle reference. All pages supporting retained method, numerical, evaluation, limitation, and proposed-check claims are included. Six crops reproduce original PDF content; narrow figures/tables were rendered at higher native DPI. The dataset-comparison crop retains its necessary explanatory footnotes. The paper contains no appendix; the separate supplement, dataset, and code remain outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract
- 1. Introduction
- 2. Related work
- 3. Zenseact Open Dataset
- 3.1. Sensor suite
- 3.2. Calibration and coordinate systems
- 3.3. Privacy protection
- 3.4. Data categories
- 3.5. Annotations
- 4. Dataset analysis
- 4.1. Diversity
- 4.2. Long-range perception
- 4.3. Long-tail perception
- 4.4. Anonymization
- 5. Conclusion
- 6. Limitations
- 7. Acknowledgements
- References
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction-only figure limitation above was resolved by visually inspecting all eleven supplied PDF pages and all six final original crops.
- No appendix is present in this PDF. No separate supplement was supplied: the detailed annotation taxonomies, image-compression experiments, and expanded anonymization results referenced by the paper remain unreviewed.
- Dataset files, the development kit, linked code, license webpage, and other cited works were not inspected; no experiments were reproduced.
- Identity/version: the title and all ten authors match the catalog, apart from title capitalization. The title page identifies the CVF open-access ICCV 2023 version and states that it matches the accepted version except for the watermark. The supplied PDF is printed as pages 20178–20188, whereas the catalog BibTeX gives 20121–20131. This bibliographic discrepancy is preserved; no separate revision number or revision date is established, and no later edition was substituted.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1 (printed p. 20178), title/author block, CVF notice, Abstract; PDF p. 11 footer (20188)
Title matches the supplied work, with sentence-style capitalization. Authors are Mina Alibeigi, William Ljungbergh, Adam Tonderski, Georg Hess, Adam Lilja, Carl Lindström, Daria Motorniuk, Junsheng Fu, Jenny Widahl, and Christoffer Petersson; affiliation Zenseact. The first three share equal contribution. The notice identifies the open-access accepted version; the abstract states two years of collection.
Go to primary source ↓e02PDF p. 2, Section 1, diversity motivation, contributions, traffic signs, and license paragraphs
The paper motivates less-correlated geographic sampling and long-range perception, describes 446k traffic-sign instances across 156 classes, and states that ZOD is released under CC BY-SA 4.0 for research and commercial use subject to the license.
Go to primary source ↓e03PDF p. 3, Table 1, ZOD/Waymo/Argoverse/nuScenes/PandaSet rows and all footnotes
Frames: 100k, 705 km², 14 countries, 245 m range, 254k average LiDAR points, 120° camera and no map. Sequences: 1473; Drives: 29. The 55.6-hour Frames figure counts two-second LiDAR snippets, not continuous camera video. Annotation range is the 99.9th percentile; PandaSet lists 300 m.
Go to primary source ↓e04PDF p. 4, Table 2, Section 3.1, and Figure 2
The suite includes VLS128 plus two VLP16 LiDARs at 10 Hz, a 3848 × 2168 front camera with 120° horizontal field of view, and OxTS RT3000 at 100 Hz. Figure 2 depicts individual sensor axes and distinguishes upward/downward markers. LiDAR samples include timestamps, coordinates, intensity, and diode indices.
Go to primary source ↓e05PDF p. 5, Section 3.2; PDF p. 4, Figure 2
The ISO-8855 reference is tied to the rear-axle center under the calibration load with X-forward, Y-left, Z-up. Sensors are synchronized and regularly calibrated; each datapoint includes intrinsic/extrinsic calibration. Figure 2 shows local sensor frames rather than a universal common axis orientation.
Go to primary source ↓e06PDF p. 5, Section 3.1 continuation and Figure 3 with caption
Combined LiDAR clouds average about 254k points. High-precision navigation records include geographic/Cartesian coordinates and motion information at 100 Hz. Camera images are 10 Hz JPG with lossless PNG also described. Sequences/Drives include vehicle controls at 100 Hz, consumer IMU at 50 Hz, and satellite positioning at 1 Hz. Figure 3 shows annotations and 200 m of recorded future ego trajectory.
Go to primary source ↓e07PDF p. 5, Section 3.3; PDF p. 6, Section 3.4, Frames and Sequences
Faces and license plates are anonymized using blurring or Deep Neural Anonymization (DNAT). Frames provide both image versions, whereas Sequences provide blurred images only. The described public images do not establish public access to original unmodified evaluation images.
Go to primary source ↓e08PDF p. 6, Section 3.4, Frames, Sequences, and Drives
Frames provide ±1 second of LiDAR and navigation from five seconds before until 25 seconds after or 300 m ahead, whichever occurs first. Metadata includes location, country, weather, solar elevation, road type, and object counts. The 1473 twenty-second Sequences have middle-keyframe annotations; 29 Drives span minutes in two cities with varied road structures and some loop closures.
Go to primary source ↓e09PDF pp. 6–7, Section 3.5
Labels are manually produced and quality checked. Segmentation has 15 top-level classes and lane-instance properties. 3D boxes require camera/LiDAR visibility and comprise center, dimensions, and quaternion rotation; the source calls them 9-DOF. Dynamic and static hierarchies and 156 traffic-sign classes are described. Exact taxonomies are deferred to the supplement.
Go to primary source ↓e10PDF p. 7, Section 4.1 and Figure 5; PDF p. 8, Section 4.1 continuation
At radius 75 m, annotated-frame diversity area is 705 km² versus 5, 17, and 76 km² for nuScenes, Argoverse 2, and Waymo. The alternative 150 m radius gives 2039 km². Figure 5 reports clear 61.2%, cloudy 19.0%, rain 15.7%, fog 2.1%, snow 2.0%; day 77.3%, night 19.0%, twilight 3.6%.
Go to primary source ↓e11PDF p. 8, Section 4.2, Table 4, and footnote 2
CenterPoint uses mAP and CDS following reference [30]. Table 4 lists mAP/CDS of 0.25/0.46 at 0–150 m, 0.33/0.62 at 0–50 m, 0.19/0.33 at 50–100 m, 0.06/0.08 at 100–150 m, and 0.01/0.02 at 150–250 m. Footnote 2 gives Argoverse 2 mAP 0.18 at 0–150 m. No split size or training configuration is specified here.
Go to primary source ↓e12PDF p. 8, Section 4.3 opening; PDF p. 9, Figure 7 and Section 4.3 continuation
CenterPoint is evaluated by class over 0–250 m. Figure 7 shows strong class disparities; the text states that nearly half the analyzed classes have AP below 10%.
Go to primary source ↓e13PDF p. 8, Table 5; PDF p. 9, Section 4.3, traffic-sign paragraph
A ResNet-50-based traffic-sign classifier excludes classes with fewer than ten validation signs. In percent, F1-macro is 78.5, F1-micro 95.4, average accuracy for the ten most common classes 93.4, and for the ten least common classes 65.4. No uncertainty or split size is given.
Go to primary source ↓e14PDF p. 8, Table 3 and caption; PDF p. 9, Section 4.4
Original/DNAT/blur training is evaluated on original images. Three-run mean ± SD AP: Faster-RCNN 30.23 ± 0.09 / 30.28 ± 0.03 / 30.17 ± 0.06; YOLOv7 33.62 ± 0.04 / 33.74 ± 0.09 / 33.67 ± 0.01. Training concerns vehicles, vulnerable vehicles, and pedestrians. The authors claim no statistically significant degradation without naming a test.
Go to primary source ↓e15PDF p. 9, Sections 5 and 6
Temporal-consistent sequence labels, supporting camera frames, more recordings, and maps are future work. Current labels do not support consistent tracking; distant 3D label quality and recall decline, other anonymization tasks remain unexplored, and some Frames lack traffic-sign or ego-road labels. Section 5 uses 245 m while Section 6 says 250 m.
Go to primary source ↓e16PDF p. 5, Section 3.1 camera paragraph; PDF pp. 6–7, Section 3.5; PDF p. 9, Section 4.4
The source directs readers to separate supplementary material for image-compression evidence, detailed annotation taxonomies, and expanded anonymization experiments. These details are not contained in the supplied eleven-page paper.
Go to primary source ↓8.5 Primary sources
Zenseact Open Dataset: A Large-Scale and Diverse Multimodal Dataset for Autonomous Driving ↗
PDF · 7,633 extracted words
Source fingerprint
2b9f56a9efe476349359bfc3c35d7d00c5a077a1bb4fca0ca44d29fed280b491