Scalability in Perception for Autonomous Driving: Waymo Open Dataset
1. Paper overview
In one sentence: Synchronized multisensor sequences and independent spatial labels make perception scale and geographic transfer measurable, while protocol ambiguities limit exact reproduction. e01-identitye02-purposee03-sensorse04-labelse10-detection-metrice15-detection-resultse18-domain-resultse19-size-results
| At a glance | What to know |
|---|---|
| Research problem | Source description Autonomous-driving perception needs geographically varied, accurately synchronized and densely labeled data. This release targets generalization within and between operating regions through 1,150 twenty-second scenes. e02-purpose |
| Core mechanism | Source description Five cameras and five LiDARs provide separate image and 3D annotations with temporal IDs. Range images preserve information beyond a point set, supporting representation and sensor-fusion research. e03-sensorse04-labelse05-range |
| A key reported result | Geographic transfer of 3D detection: Vehicles: SF→SUB 36.7 versus SUB→SUB 44.0; SUB→SF 42.5 versus SF→SF 49.2. APH. LEVEL 2 validation; SF versus SUB (Phoenix plus Mountain View); vehicle/pedestrian IoU 0.7/0.5. Reader-computed losses are 7.3 and 6.7 points; prose instead says 8.0 and 7.6. Table 9 supports transfer loss, but class exposure is imbalanced and prose/table magnitudes conflict. e17-domain-protocole18-domain-results |
| Reading caution | Reader analysis Representativeness is bounded by the collected geographies and conditions: 646 of 798 training scenes are daytime. This release provides no maps or closed-loop planning benchmark. e07-splitse20-scope-reproduction |
Core contributions
- Source description
Five cameras and five LiDARs provide separate image and 3D annotations with temporal IDs. Range images preserve information beyond a point set, supporting representation and sensor-fusion research. e03-sensorse04-labelse05-range
- Source description
Fixed 798/202/150 train/validation/test splits support detection and tracking. The test set is geographically held out; domain and cumulative-data experiments probe generalization and scale. e02-purposee07-splitse17-domain-protocole19-size-results
Figure 1. Sensor-specific axes make calibration a prerequisite for combining observations. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the vehicle axes on the roof: red x points forward, green y points left, and the blue symbol denotes upward z. Then inspect the local arrows beside each LiDAR and within the enlarged camera arrangement. Their differing orientations matter: a point expressed in a sensor frame cannot be treated directly as a vehicle-frame point. Section 3.2 specifies sensor-to-vehicle extrinsics and a separate vehicle-to-global pose transform. The red camera wedges indicate viewing directions; they are not data-flow arrows or neural-network branches. This diagram describes the acquisition geometry underlying the separate camera and LiDAR annotations. e03-sensorse04-labelse12-pointpillars
What it supports. The dataset exposes the transformations needed to relate multiple observations to a common moving vehicle frame. That supports sensor-fusion research, but the baseline 3D detector evaluated here uses LiDAR alone. The availability of aligned modalities therefore establishes an experimental opportunity, not a measured fusion gain.
Where the evidence stops. Figure 1 calls the back LiDAR REAR, while Table 2 expands R as Right. The crop preserves that naming discrepancy. Section 3.2’s transform definitions, rather than inferred sensor names, govern the geometry.
2. Motivation
2.1 The problem and the proposed response
Autonomous-driving perception needs geographically varied, accurately synchronized and densely labeled data. This release targets generalization within and between operating regions through 1,150 twenty-second scenes. e02-purpose
2.2 What this reading follows
This paper turns an expensive data-collection problem into a shared perception benchmark. Its contribution runs from sensor geometry and capture timing to separate image and LiDAR labels, stable track identities and predefined evaluation splits. The six visuals below follow that chain, then examine what the baselines reveal about range, tracking errors, geographic transfer and training-set size. Read every table’s header carefully: a slash can mean difficulty levels in one table and AP versus APH in another. The supplied CVPR 2020 version also contains conflicting metric, numerical and split descriptions. Those conflicts remain visible here because they affect how a researcher would reproduce the results. e01-identitye02-purposee03-sensorse04-labelse10-detection-metrice15-detection-resultse18-domain-resultse19-size-results
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 multisensor collection, spatial/temporal labels and perception protocols. Architecture, prediction paradigm and quadrant are not applicable: the detector/tracker baselines do not establish a world-action model or joint future/action prediction. e02-purposee04-labelse12-pointpillarse14-trackere20-scope-reproduction
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 |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 From a moving sensor to a supervised perception example
Begin with a return expressed in its LiDAR’s coordinates. The sensor extrinsics map it into vehicle coordinates, and the vehicle pose relates that frame to the global frame. Acquisition time matters because the vehicle moves while sensors scan. The release therefore supplies per-return pose information and rolling-shutter-aware camera projections; its projection procedure estimates capture time under constant linear and angular velocity. The independent labels are another important distinction: image boxes tightly bound visible objects, whereas LiDAR boxes describe upright 3D geometry. Projecting the latter produces an amodal image target, not the same annotation. PointPillars then takes a particular route through this richer resource: single-frame LiDAR points become a BEV representation and detection proposals. The benchmark permits broader inputs, but the displayed detector does not measure camera–LiDAR fusion gains. e03-sensorse04-labelse05-rangee06-projectione10-detection-metrice12-pointpillarse20-scope-reproduction
Figure 4. The released LiDAR representation preserves separate measurements and returns. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the six bands from top to bottom using the original caption’s ordering: range, intensity and elongation from the first return, followed by those channels from the second return. Each band depicts the front 90 degrees, rather than the full sensor panorama. In the underlying range-image representation, columns correspond to azimuth and rows to inclination; Section 3.4 says the corrected center column points along positive x and row zero has maximum inclination. Compare the visible scene structure across channels and returns, but do not interpret the grayscale as a shared numerical scale: the figure provides no color bar. e05-rangee12-pointpillarse20-scope-reproduction
What it supports. A range image preserves the sensor’s sampling arrangement as well as measurement channels and two returns. The visibly different bands explain why reducing the release to an unstructured set of coordinates can omit useful information. The paper motivates alternative LiDAR representations without demonstrating their superiority in this figure.
Where the evidence stops. This is one qualitative example with no channel-wise detection ablation. The sparse-looking second-return bands do not quantify information value, and the figure cannot establish gains from elongation or from using both returns.
5.2 Separate localization, orientation and identity errors
The detection and tracking tables measure different parts of the perception problem. AP evaluates scored detections under box matching; APH is intended to add heading sensitivity. MOTA instead accumulates misses, false positives and identity mismatches over time, while the paper’s MOTP is a matched-box distance. A reader should therefore ask which component changed before calling one result better. For example, the pedestrian AP–APH difference signals a reported orientation-sensitive deficit, whereas Table 7’s large miss percentages point to detection/localization opportunities. These are interpretations of the reported scores, not isolated causal experiments. There is also a prerequisite for trusting a new APH implementation: the printed heading factor equals zero for a perfect heading and one for an opposite heading. Its description as accuracy is inconsistent, so exact reproduction requires resolving that semantic conflict rather than quietly substituting a preferred formula. e10-detection-metrice11-tracking-metrice15-detection-resultse16-tracking-results
5.3 Ask what more data changes before attributing a gain
Table 10 asks what happens as cumulative training subsets grow, whereas Table 9 asks how training and evaluation geography interact. Together they motivate a more careful question than whether a dataset is large. More sequences can change object exposure, urban/suburban mixture and optimization effort at once. Table 8 makes one confound concrete: SF has 2.0 million training pedestrian objects against 210 thousand in SUB. The SF pedestrian validation result also favors SF-only training over ALL in Table 9. Reader interpretation: geographic relevance and total sample count can pull in different directions for this baseline. A controlled follow-up would match exposure and compute before interpreting the residual transfer gap. The original tables establish useful observations, but their unbalanced domains, inconsistent prose and unresolved Table 10 split do not establish a universal relationship between coverage and generalization. e17-domain-protocole18-domain-resultse19-size-resultse20-scope-reproduction
5.4 Training and inference
During training
Both PointPillars class models use first-block stride one. Anchors are (4.73,2.08,1.77) m for vehicles and (0.9,0.86,1.71) m for pedestrians, at headings 0 and π/2. Wrapped heading residuals use smooth-L1 with Huber delta 1/9. e12-pointpillars
Faster R-CNN/ResNet-101 is COCO-pretrained then fine-tuned. The tracker called Tracktor also uses a fine-tuned COCO-pretrained Faster R-CNN. Frozen-module policies and complete training schedules are unspecified. e13-camera-baselinese20-scope-reproduction
During inference
The 3D tracker reuses trained detections, rejects scores below 0.2, associates by Hungarian matching with 1−IoU cost and minimum IoU 0.5, then updates a constant-velocity Kalman state containing box geometry, heading and velocity. Track scores gain matched confidence or lose 0.3, clipped to [0,3]. It estimates object motion without executing vehicle actions. e14-trackere20-scope-reproduction
Camera detection runs on each of five images and aggregates evaluation. The 2D task permits a single camera image per prediction; the 3D task permits multimodal and preceding-frame inputs, although this baseline uses single-frame LiDAR. e13-camera-baselinese10-detection-metrice12-pointpillars
5.5 Implementation flow
- Preserve geometry and capture time
Extrinsics map sensors into the vehicle frame, whose axes point forward, left and up; vehicle pose maps into the global frame. Two-return range images retain intensity, elongation and acquisition-time pose. Rolling-shutter projection estimates capture time under constant linear/angular velocity. e03-sensorse05-rangee06-projection
- Keep annotation modalities distinct
LiDAR labels cover vehicles, pedestrians, signs and cyclists with upright 7-DOF boxes. Independent image boxes cover vehicles, pedestrians and cyclists. Tight image labels differ from amodal projections. LEVEL 2 includes LEVEL 1; single-frame 3D evaluation excludes zero-point labels and treats hard or at-most-five-point labels as harder examples. e04-labelse10-detection-metric
- Encode and detect
PointPillars uses single-frame data from all LiDARs, single-layer PointNet encoding, a BEV pseudo-image and a CNN proposal network. The stated grid is 512×512 at 0.33 m resolution, with x/y extent ±85 m and z extent ±3 m. e12-pointpillars
- Evaluate predictions and identities
Detection uses Hungarian assignment across score thresholds. Vehicle/pedestrian IoU thresholds are 0.7/0.5. Tracking reports maximum MOTA over sampled cutoffs; MOTP uses 1−IoU, so lower means smaller localization error. e10-detection-metrice11-tracking-metrice15-detection-results
6. Experiments & results
Waymo Open Dataset supplies synchronized camera/LiDAR sequences, independent spatial annotations and perception benchmarks. Its PointPillars results expose long-range difficulty, geographic transfer loss and gains from more training sequences. The evidence concerns offline detection and tracking; printed metric and protocol inconsistencies constrain exact reproduction.
This dataset paper has no neural-network architecture figure and no controlled sensor-feature or fusion-benefit ablation. Figures 1 and 4 explain sensing and representation; Tables 9 and 10 supply geographic and training-size diagnostics rather than component ablations. The edition preserves those experiment types without inventing a model diagram or a fusion result. e03-sensorse05-rangee12-pointpillarse18-domain-resultse19-size-resultse20-scope-reproduction
6.1 Read the original evidence
Table 6. Detection quality depends strongly on distance and on what the metric measures. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First choose BEV or full 3D evaluation; these are different box-overlap problems. Next select AP or APH and keep the slash ordering fixed: every entry is LEVEL 1 followed by cumulative LEVEL 2. In the right-hand 3D group, follow the vehicle APH row from Overall through the three distance bins. Section 5.1 identifies this as the hidden 150-scene test set, using vehicle IoU 0.7 and pedestrian IoU 0.5. The single-frame PointPillars baseline consumes all LiDARs. Do not mix the left BEV overall score with the right 3D score when comparing difficulty or heading effects. e10-detection-metrice12-pointpillarse15-detection-results
What it supports. Vehicle LEVEL 2 3D APH falls from 80.8 at 0–30 m to 52.3 at 30–50 m and 26.7 in the far bin. Pedestrian overall LEVEL 2 AP is 55.9 while APH is 45.1. These are distinct reported challenges in distant detection and heading-sensitive evaluation.
Where the evidence stops. The printed APH weight in Section 4.1.1 increases with heading error, conflicting with its description as accuracy. Scores are transcribed as reported. This table supplies no uncertainty estimates or controlled explanation for the distance effect.
Table 7. The tracking error decomposition points primarily to misses. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read Overall MOTA before inspecting its error components. Each slash again separates LEVEL 1 from LEVEL 2. Miss, Mismatch and FP explain the errors subtracted in the MOTA definition; they are not alternative measures of box overlap. MOTP is different: the paper defines it with distance equal to one minus IoU, so smaller values mean smaller matched-box localization error. The right group reports MOTA by range. Section 5.2 links these results to PointPillars detections, Hungarian assignment and a constant-velocity Kalman filter. Association uses a minimum IoU of 0.5, distinct from the vehicle evaluation threshold of 0.7. e11-tracking-metrice14-trackere15-detection-resultse16-tracking-results
What it supports. At LEVEL 2, vehicle MOTA is 40.1 and pedestrian MOTA is 37.7. Their miss percentages, 43.4 and 50.2, greatly exceed identity-mismatch percentages of 0.13 and 0.47. This supports examining detection and localization failures before attributing the whole tracking deficit to identity assignment.
Where the evidence stops. Table 7 and the tracking paragraph do not explicitly identify the evaluation split. A low mismatch percentage does not prove that association is solved; missed detections can remove difficult opportunities for identity assignment.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Vehicle 3D detection Single-frame PointPillars; hidden 150-scene test set; IoU 0.7; LEVEL 1/2. | AP 63.3/55.6; APH 62.8/55.1. 3D AP / APH, 0–100 scale | LEVEL 2 APH: 80.8 at 0–30 m, 52.3 at 30–50 m, 26.7 beyond 50 m. Reported performance declines sharply with range; this does not isolate point sparsity from other range-correlated factors. e12-pointpillarse15-detection-results |
| Pedestrian 3D detection Same hidden test set; IoU 0.5; LEVEL 1/2. | AP 62.1/55.9; APH 50.2/45.1. 3D AP / APH, 0–100 scale | LEVEL 2 AP exceeds APH by 10.8 points, computed from Table 6. Heading-sensitive scores are substantially lower; the printed heading-weight inconsistency remains unresolved. e10-detection-metrice15-detection-results |
| Camera-only 2D detection Faster R-CNN/ResNet-101; five-camera aggregation; LEVEL 1/2; split not explicitly stated in the result paragraph. | Vehicles 63.7/53.3; pedestrians 55.8/52.7. AP | One supplied camera-detector baseline; no fusion comparison. Tight image-box results cannot establish 3D localization. e04-labelse13-camera-baselines |
| Online 3D multi-object tracking PointPillars plus Hungarian/Kalman tracking; LEVEL 1/2; Table 7 does not explicitly identify the split. | Vehicles 42.5/40.1; pedestrians 38.9/37.7. MOTA | LEVEL 2 miss/mismatch percentages: vehicles 43.4/0.13; pedestrians 50.2/0.47. Misses dominate identity mismatches; underlying localization/recall causes are not isolated. e11-tracking-metrice14-trackere16-tracking-results |
| Camera-only vehicle tracking Tracker named Tracktor; LEVEL 1/2; split not explicitly stated. | 34.8/28.3. MOTA | No additional 2D tracking comparator reported. The baseline name and reference [14] disagree, limiting implementation identification. e13-camera-baselinese21-tracking-citation |
| Geographic transfer of 3D detection LEVEL 2 validation; SF versus SUB (Phoenix plus Mountain View); vehicle/pedestrian IoU 0.7/0.5. | Vehicles: SF→SUB 36.7 versus SUB→SUB 44.0; SUB→SF 42.5 versus SF→SF 49.2. APH | Reader-computed losses are 7.3 and 6.7 points; prose instead says 8.0 and 7.6. Table 9 supports transfer loss, but class exposure is imbalanced and prose/table magnitudes conflict. e17-domain-protocole18-domain-results |
| Training-set size sensitivity Cumulative 10/30/50/100% training-sequence subsets; LEVEL 2; Table 10 says validation, Section 5.4 says test. | Vehicles: 29.7/28.9 → 49.8/49.4; pedestrians: 39.5/27.7 → 53.0/43.0, from 10% to 100%. AP/APH | Vehicle AP/APH: 41.4/41.0 at 30%, 46.3/45.8 at 50%. Scores improve across these subsets. Split ambiguity, missing compute controls and absent uncertainty preclude a precise scaling-law claim. e19-size-resultse20-scope-reproduction |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 9. Training geography changes validation performance, with unequal object exposure. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The arrow points from training domain to validation domain. SUB combines Phoenix and Mountain View; SF is San Francisco; ALL combines them. Crucially, the ordering before the arrow changes: the left column is ALL/SUB/SF, while the right is ALL/SF/SUB. Every number is LEVEL 2 APH, with vehicle IoU 0.7 and pedestrian IoU 0.5 as stated in the caption. Compare the second and third values within each vehicle cell to contrast same-domain and cross-domain training at a fixed evaluation domain. Table 8 supplies a necessary control warning: the domains contain substantially different numbers of labeled objects. e17-domain-protocole18-domain-results
What it supports. The table implies vehicle transfer losses of 7.3 points on SUB and 6.7 on SF, computed from 44.0−36.7 and 49.2−42.5. Pedestrian SF evaluation favors SF-only training, 47.6, over ALL, 46.0. More total data does not uniformly improve this particular geographic evaluation.
Where the evidence stops. The prose reports vehicle losses of 8.0/7.6 and SF pedestrian differences of 2.4/19.8; the table yields 7.3/6.7 and 1.6/17.9. Preserve both accounts. Unequal object counts also prevent interpreting the comparison as geography alone.
Table 10. Larger cumulative training subsets improve both reported detection scores. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across each row as the fraction of training sequences grows from 10% to 100%. Unlike Tables 6 and 7, the slash here separates AP from APH, not difficulty levels: the entire table is LEVEL 2. The subsets are cumulative, so a larger subset contains the smaller one. Section 5.4 uses the same PointPillars baseline and describes the gains without requiring augmentation. Keep this experiment separate from Table 6: its caption names validation, while its surrounding paragraph names test. That unresolved split conflict rules out treating the 100% column as a straightforward replication of the main test table. e12-pointpillarse19-size-resultse20-scope-reproduction
What it supports. Vehicle APH rises from 28.9 to 49.4, and pedestrian APH from 27.7 to 43.0, across the displayed training fractions. Both rows improve monotonically at these sampled budgets. The result supports a data-volume benefit for this configuration, without establishing a general scaling law.
Where the evidence stops. The evaluation split is internally inconsistent, and training compute, subset seeds and repeated-run uncertainty are not supplied. Cumulative sampling reduces one source of variation but cannot isolate sample diversity from extra optimization or geographic composition.
7. Analysis & limitations
7.1 What the evidence leaves open
Representativeness is bounded by the collected geographies and conditions: 646 of 798 training scenes are daytime. This release provides no maps or closed-loop planning benchmark. e07-splitse20-scope-reproduction
Counts and coverage definitions conflict: Table 1 lists 9.9M 2D boxes versus roughly 12M in Section 3.5/Table 5. Its 75 m dilation radius differs from the 150 m wording elsewhere. Treat reported 76 km² coverage as definition-dependent. e08-countse09-geography
7.2 Questions for discussion
- After matching pedestrian exposure and training compute, how much SF–SUB transfer loss remains?
- Which metric implementation and evaluation split generated the APH and dataset-size tables?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs matching release/splits, calibration and timing, annotation filters, evaluation code and baseline configurations. Optimizer, schedule, batch size, hardware and complete track-lifecycle thresholds are omitted; the PDF gives no license terms. Public code/data availability is an author statement, not independently verified here. e03-sensorse10-detection-metrice14-trackere20-scope-reproduction
Proposed checks should resolve APH semantics and the Table 10 split, then control object exposure in geographic comparisons. Assuming a later implementation matches this paper cannot resolve those ambiguities. e10-detection-metrice17-domain-protocole19-size-results
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: Resolve the heading metric with controlled synthetic predictions
Reader-proposed check, not performed: once the version-matched evaluator is available, construct one ground-truth box and one prediction with identical center, dimensions and score. Test equal headings, an intermediate error and opposite headings while holding the IoU match fixed; also test equivalent headings on either side of the ±π boundary. Record AP, the true-positive heading weight and APH separately. The printed factor predicts zero weight for equal headings and one for opposite headings. An evaluator intended to reward accuracy should show the reverse ordering without changing AP. This falsifiable comparison would identify whether the equation, its description or the implementation must be treated as the unresolved source of the mismatch; document the result without silently changing the published table. e10-detection-metrice15-detection-results
Check 2: Test geographic transfer under matched pedestrian exposure
Reader-proposed check, not performed: reproduce the SF/SUB filtering on the original training and validation memberships. Train the Section 5.1 pedestrian baseline on SF and SUB subsets matched for labeled pedestrian exposure, with fixed architecture, optimizer settings and training-update budget. Evaluate both models on the same SF validation set at IoU 0.5 and LEVEL 2, repeating sequence sampling across seeds. Keep the full-data models as reference controls and report AP as well as resolved APH. Persistence of a large transfer gap after matching exposure would support a geographic effect beyond raw object count; a much smaller gap would implicate that count imbalance. Separately label any Table 10 comparison with the verified split, because its caption and prose disagree. e10-detection-metrice12-pointpillarse17-domain-protocole18-domain-resultse19-size-results
8.3 Reading coverage
Visual audit: All nine supplied PDF pages were rendered at 200 DPI and visually inspected, including the title/author page, all six figures, all ten tables, metric equations, detector/tracker configuration paragraphs and the references. Six final original crops were separately viewed: Figure 1 at 400 DPI, Figure 4 at 350 DPI, Tables 6–7 at 200 DPI, and Tables 9–10 at 400 DPI. Figure 1 axes and labels were checked against Section 3.2/Table 2; Figure 4 channel order against its caption; table headings and arrows against Sections 4–5. Inspection confirmed the printed APH-weight conflict, Table 9 prose/table differences, Table 10 split conflict and Tracktor/reference mismatch. The supplied PDF has no appendix. Separate supplements, dataset files, code and later releases remain outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9. 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
- 3.1. Sensor Specifications
- 3.2. Coordinate Systems
- 3.3. Ground Truth Labels
- 3.4. Sensor Data
- 3.5. Dataset Analysis
- 4.1. Object Detection
- 4.2. Object Tracking
- 5.1. Baselines for Object Detection
- 5.2. Baselines for Multi-Object Tracking
- 5.3. Domain Gap
- 5.4. Dataset Size
- 6. Conclusion
- 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-image limitation above was addressed by visually inspecting all nine PDF pages and all six final crops.
- No separate supplement was supplied; the PDF contains no appendix. Code and dataset files were not inspected, and no experiments were reproduced.
- Edition scope: the CVPR 2020 CVF open-access title page says this version is identical to the accepted version except for its watermark. The IEEE published artifact and later dataset revisions were not supplied or compared. All 23 catalog authors match; the title page spells Aurélien with an accent.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01-identityPDF p. 1, title block, affiliation footnote and CVF watermark
The matching title and 23-author list identify this work. The title page spells Aurélien Chouard with an accent and lists Waymo LLC and Google LLC. Yu Zhang’s footnote says work done while at Waymo LLC. The CVPR 2020 open-access watermark says the artifact is identical to the accepted version except for the watermark.
Go to primary source ↓e02-purposePDF pp. 1–2, Abstract and Section 1; p. 5, Section 4
The release contains 1,150 twenty-second scenes with synchronized camera/LiDAR data, separate spatial annotations and temporal identifiers. Test scenes are geographically held out. The defined tasks are 2D/3D detection and tracking.
Go to primary source ↓e03-sensorsPDF p. 2, Section 3.1 and Table 2; p. 3, Table 3, Figure 1 and Section 3.2
Five LiDARs and five cameras are used. Top LiDAR data are restricted to 75 m and other LiDARs to 20 m, with two returns per shot. Vehicle axes are forward/left/up; extrinsics map sensor to vehicle coordinates and vehicle pose maps vehicle to global coordinates. Figure 1 labels the back sensor REAR, whereas Table 2 expands R as Right.
Go to primary source ↓e04-labelsPDF pp. 3–4, Section 3.3 and Figure 2; p. 5, Section 4
LiDAR annotations are upright 7-DOF boxes for vehicles, pedestrians, signs and cyclists. Camera annotations are independently drawn tight axis-aligned 2D boxes for vehicles, pedestrians and cyclists. IDs persist across frames. LEVEL 2 metrics include LEVEL 1. The 3D tuple uses theta but its explanatory prose calls the heading alpha.
Go to primary source ↓e05-rangePDF p. 4, Section 3.4, Figures 3–4 and Eq. (4)
Range images retain two returns, with range, intensity, elongation, no-label-zone indicators, acquisition-time vehicle pose and camera projections. Row zero has maximum inclination; the corrected center column points along positive x. Figure 4 shows the front 90 degrees: range/intensity/elongation for the first return, followed by the same channels for the second. Figure 3 shows camera-specific timing distributions.
Go to primary source ↓e06-projectionPDF pp. 4–5, Section 3.4, rolling-shutter projection paragraph and Figure 5
Rolling-shutter projection estimates vehicle pose at an unknown point-capture time under constant linear/angular velocity, projects into the image and minimizes disagreement with the pixel’s capture time. Figure 5 is a qualitative overlay.
Go to primary source ↓e07-splitsPDF p. 5, Table 4 and Section 4; p. 6, Section 4.2
There are 798 training, 202 validation and 150 test scenes. Training counts are PHX 286, MTV 103, SF 409; day 646, night 79, dawn 73. Tracking sequences last 20 seconds with 10 Hz sensor sampling.
Go to primary source ↓e08-countsPDF p. 2, Table 1, 2D Boxes row; p. 5, Section 3.5 and Table 5
Table 1 lists 9.9M 2D boxes; Section 3.5 says around 12M and Table 5 lists 9.0M vehicle, 2.7M pedestrian and 81k cyclist image objects. The PDF does not reconcile these counts.
Go to primary source ↓e09-geographyPDF p. 2, Table 1 and caption; p. 5, Section 3.5; p. 6, Figure 6; p. 8, Section 6
The reported visited area is 76 square kilometres. Table 1 specifies a 75 m dilation radius; Section 3.5 says 150-meter-diluted ego poses and the conclusion refers to 150 m visibility. Figure 6 depicts covers of touched level-13 S2 cells. The dilation wording is not reconciled.
Go to primary source ↓e10-detection-metricPDF pp. 5–6, Section 4.1.1, Eqs. (5)–(6); p. 7, Section 5.1 difficulty definition
The 3D task permits any camera/LiDAR sensors and preceding-frame inputs. AP and APH integrate interpolated precision over recall. The printed APH weight is normalized wrapped heading difference: zero for equal headings and one for opposite headings, despite being called accuracy. Matching maximizes total IoU with the Hungarian algorithm over sampled score thresholds. Zero-point labels are ignored; hard labels or labels with at most five points enter the harder group, with cumulative LEVEL 2 reporting.
Go to primary source ↓e11-tracking-metricPDF p. 6, Section 4.2, Eqs. (7)–(8)
MOTA subtracts normalized misses, false positives and mismatches from 100. MOTP is 100 times mean matched-box distance, defined as 1 minus IoU. The final tracking score is maximum MOTA over sampled score cutoffs.
Go to primary source ↓e12-pointpillarsPDF pp. 6–7, Section 5.1, 3D LiDAR Detection
The single-frame baseline uses all LiDARs, single-layer PointNet encoding into BEV and a CNN proposal network. Voxel size is 0.33 m, x/y extent ±85 m and z extent ±3 m; the stated grid is 512×512. Both class models use first-block stride one. Vehicle anchors are (4.73,2.08,1.77) m and pedestrian anchors (0.9,0.86,1.71) m, at headings 0 and pi/2. Wrapped heading residuals use smooth-L1 with Huber delta 1/9.
Go to primary source ↓e13-camera-baselinesPDF p. 6, Section 4.1.2; p. 7, Sections 5.1–5.2, 2D detection and tracking paragraphs
The 2D task uses a single camera image without LiDAR. Faster R-CNN/ResNet-101 is COCO-pretrained then fine-tuned; predictions from five cameras are aggregated. Vehicle AP is 63.7/53.3 and pedestrian AP 55.8/52.7 at LEVEL 1/2. The tracker called Tracktor uses fine-tuned COCO-pretrained Faster R-CNN, with sigma_active 0.4, lambda_active 0.6 and lambda_new 0.3; vehicle MOTA is 34.8/28.3. These result paragraphs do not explicitly identify the split.
Go to primary source ↓e14-trackerPDF p. 7, Section 5.2, 3D Tracking
The online tracker uses pretrained PointPillars detections, 1−IoU cost, Hungarian assignment and constant-velocity Kalman filtering. It drops scores below 0.2 and requires association IoU at least 0.5. State contains center, dimensions, heading and three velocity components. Track scores gain matched detection confidence, lose 0.3 if unmatched and are clipped to [0,3]. Numerical Pending/Live/Dead thresholds are omitted.
Go to primary source ↓e15-detection-resultsPDF p. 7, Section 5.1 evaluation paragraph; p. 8, Table 6, 3D columns
Table 6 evaluates the hidden 150-scene test set at vehicle/pedestrian IoU 0.7/0.5. Vehicle APH is 62.8/55.1 and AP 63.3/55.6 at LEVEL 1/2; pedestrian APH is 50.2/45.1 and AP 62.1/55.9. Vehicle LEVEL 2 APH is 80.8 at 0–30 m, 52.3 at 30–50 m and 26.7 at 50 m and beyond.
Go to primary source ↓e16-tracking-resultsPDF p. 8, Table 7; p. 7, Section 5.2 interpretation
Vehicle 3D MOTA is 42.5/40.1 and pedestrian MOTA 38.9/37.7 at LEVEL 1/2. Vehicle misses are 40.0/43.4, mismatches 0.14/0.13 and false positives 17.3/16.4; pedestrian misses are 48.6/50.2, mismatches 0.49/0.47 and false positives 12.0/11.6. The caption and tracking paragraph do not explicitly state the split.
Go to primary source ↓e17-domain-protocolPDF pp. 7–8, Section 5.3; p. 8, Table 8
SF is San Francisco, SUB combines Mountain View and Phoenix, and ALL combines the domains. Filtering preserves training/validation membership. Table 8 reports 2.0M SF versus 210K SUB training pedestrian objects and 2.9M versus 1.9M vehicle objects; object exposure is unbalanced.
Go to primary source ↓e18-domain-resultsPDF p. 8, Table 9 and Section 5.3 continuation; p. 7, final paragraph
Table 9 reports LEVEL 2 validation APH at vehicle/pedestrian IoU 0.7/0.5. ALL/SUB/SF→SUB: vehicles 45.3/44.0/36.7, pedestrians 25.7/20.6/19.9. ALL/SF/SUB→SF: vehicles 50.3/49.2/42.5, pedestrians 46.0/47.6/29.7. Vehicle losses computed from the table are 7.3 and 6.7 points, versus 8.0 and 7.6 in prose. SF pedestrian differences are 1.6 and 17.9 points, versus 2.4 and 19.8 in prose.
Go to primary source ↓e19-size-resultsPDF p. 8, Section 5.4 and Table 10 including caption
PointPillars is trained on cumulative random sequence subsets of 10%, 30%, 50% and 100%, described without requiring augmentation. Vehicle LEVEL 2 AP/APH: 29.7/28.9, 41.4/41.0, 46.3/45.8, 49.8/49.4; pedestrian: 39.5/27.7, 45.7/35.7, 50.3/40.4, 53.0/43.0. The caption says validation but the paragraph says test. No repeated-seed statistics or error bars are given.
Go to primary source ↓e20-scope-reproductionPDF p. 2, Table 1 Maps row; pp. 5–8, Sections 4–6
Table 1 reports no maps. Section 6 treats maps, more weather/behavior diversity, prediction and planning as future work and claims public data/code availability. Sections 5.1–5.4 omit full optimizer/schedule/batch/compute specifications; no license terms appear in this PDF. No closed-loop driving experiment or fusion-benefit ablation is reported.
Go to primary source ↓e21-tracking-citationPDF p. 7, Section 5.2, 2D Tracking; p. 9, reference [14]
The prose names Tracktor and cites [14], but [14] is titled Multi-object tracking with neural gating using bilinear lstm. The supplied PDF does not resolve the implementation-reference mismatch.
Go to primary source ↓8.5 Primary sources
Scalability in Perception for Autonomous Driving: Waymo Open Dataset ↗
PDF · 6,695 extracted words
Source fingerprint
69bc8edbf906fc5f86da9feae397fb6eff060142af36b1d99fcc301aca292108