BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning
1. Paper overview
In one sentence: BDD100K shows that plentiful simpler annotations can improve scarce complex perception tasks, but transfer depends on task pairing, data scale, and the error being measured. task-overviewannotation-semanticstracking-datahomogeneous-transfersemantic-transfermots-transfer
| At a glance | What to know |
|---|---|
| Research problem | Source description Driving systems need image labels, object regions, dense masks, and temporal identities together, yet expensive annotations limit data diversity. BDD100K makes annotation allocation and heterogeneous multitask learning experimentally accessible through one driving-video collection. task-overview |
| Core mechanism | Source description The benchmark covers image tagging, lane detection, drivable-area segmentation, object detection, semantic segmentation, instance segmentation, object tracking, segmentation tracking, domain adaptation, and imitation learning. These are supported task directions, not ten demonstrated outputs of one deployed model. task-overviewimitation |
| A key reported result | Multiple object tracking and segmentation: All four tasks: 23.3 AP, 41.4 MOTSA, 81.6 MOTSP; 5132/6228/472 errors. Instance AP; MOTSA; MOTSP; FN/FP/IDSW. BDD100K MOTS; 12K training frames from 60 videos plus auxiliary detection, MOT, and instance labels; evaluation split not explicitly identified. MOTS-only: 13.0 AP, 30.4 MOTSA, 81.8 MOTSP; 8352/5116/566 errors. Aggregate accuracy rises; false positives rise and MOTSP falls slightly. The 11.0 MOTSA-point gain is a calculated difference, not a controlled estimate of diversity alone. mots-transfer |
| Reading caution | Reader analysis Geographic variety is concentrated in populated US regions, and detection categories have a long tail. Diversity within this collection does not establish worldwide coverage or balanced recognition quality. collectionannotation-semantics |
Core contributions
- Source description
The benchmark covers image tagging, lane detection, drivable-area segmentation, object detection, semantic segmentation, instance segmentation, object tracking, segmentation tracking, domain adaptation, and imitation learning. These are supported task directions, not ten demonstrated outputs of one deployed model. task-overviewimitation
- Source description
Separate homogeneous, cascaded, and heterogeneous experiments expose when sharing supervision helps or hurts. The dataset contribution includes geographically and environmentally varied videos with annotations at multiple granularities. task-overviewcollectionhomogeneous-transfermots-transfer
Figure 1. Different annotation structures turn the same video collection into several learning problems. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the first two rows: textual scene tags summarize an image, rectangles localize objects, and line or area overlays describe the road. The rightmost panels add dense semantic and instance information. Then follow the bottom row horizontally, where object masks appear across a sequence rather than as isolated images. This progression explains the paper's use of heterogeneous: the output can be a whole-image category, a spatial region, a pixel label, or an identity maintained through time. The image is an overview of annotation types. Training architectures and transfer procedures are described separately in Section 5. task-overviewsegmentation-datatracking-datahomogeneous-transfermots-transfer
What it supports. The dataset supports joint study of perception tasks with different output structures. The temporal examples are especially relevant to MOTS, where a useful output must identify an object spatially and associate it over time. That additional annotation demand motivates transferring supervision from larger, simpler-label subsets.
Where the evidence stops. These are annotation examples, not evidence that one network solves all tasks or that every video has every label. The pixel and tracking subsets are substantially smaller than the full collection; the montage does not quantify their prevalence.
2. Motivation
2.1 The problem and the proposed response
Driving systems need image labels, object regions, dense masks, and temporal identities together, yet expensive annotations limit data diversity. BDD100K makes annotation allocation and heterogeneous multitask learning experimentally accessible through one driving-video collection. task-overview
2.2 What this reading follows
A driving dataset can contain many frames yet offer little supervision for the hardest tasks. BDD100K addresses this mismatch by combining a broad video collection with smaller, more expensive annotation sets. Its central question is how to use those unequal resources together. Read the paper from labels to transfer: first distinguish lane boundaries, drivable regions, object masks, and temporal identities; then compare the experiments that share their supervision. Detection labels often help downstream tasks, while apparently compatible dense labels can hurt overall segmentation. The final tracking-and-segmentation result makes the tradeoff concrete: aggregate accuracy improves, but some error types worsen. task-overviewannotation-semanticstracking-datahomogeneous-transfersemantic-transfermots-transfer
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 recorded dataset classification fits the contribution: driving perception annotations and evaluation protocols. Architecture, prediction paradigm, and world/action quadrant are not applicable to the resource itself. Shared backbones and multitask baselines do not establish a unified future-world/action model. task-overviewimitationmots-transfer
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 Start with the label budget, then define the prediction
BDD100K's basic unit is a video, but its supervision is not uniform across videos or tasks. Section 3 samples a reference image for image tasks, while tracking requires annotations over time. Pixel-level masks are costly enough to occupy a smaller subset, and MOTS narrows the training pool further to sixty videos. This organization matters before choosing a model: a box supplies location, an instance mask adds shape, and a track adds temporal association. None can be assumed to provide all the others. The benchmark asks whether abundant supervision for an easier component can compensate for scarce labels on the combined task. Reader interpretation: the transferable resource is task-relevant information across examples, rather than the raw frame count alone. collectionsegmentation-datatracking-datamots-transfer
Figure 6. Drivable-area labels express the driver's relationship to the road, beyond painted boundaries. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the red and blue overlays using the Figure 6 caption and Section 3.4: red marks the directly drivable region, while blue marks an alternative region. The latter can become available through a lane change. Compare the road context across the four examples, including daylight and night and locations with different lane visibility. The annotation task asks for more than recognizing pavement: the current driving position, surrounding vehicles, and scene context determine the functional distinction. This is why the paper treats drivable-area segmentation separately from detecting lane markings, even though both produce spatially organized outputs. annotation-semanticshomogeneous-transfer
What it supports. Direct and alternative drivable areas are different prediction targets, despite their potentially similar appearance. A model can learn accurate lane geometry without resolving those functional labels. Conversely, area supervision could help locate lane-related structure; Table 5 tests whether that relationship produces a measurable transfer benefit.
Where the evidence stops. The colored regions are supplied annotations, not tested motion plans or guarantees of safe passage. Their interpretation comes from the caption and label definition; this visual alone does not establish that a learned model understands right of way.
5.2 Separate shared training from the predictions used at evaluation
The homogeneous baseline makes sharing concrete. A DLA-34 base supports dense heads; three lane heads predict direction, continuity, and category, while drivable-area prediction is segmentation. Training adds weighted lane cross-entropy, with foreground weight ten. At evaluation, low-resolution maps are upsampled, lane responses undergo nonmaximum suppression, and the contour metric uses thinning and a specified tolerance. These processing choices are part of the measured result, not incidental display settings. Table 5 then asks whether adding another task improves the target task under a fixed image-count condition. The improvement for lane prediction at 10K does not recur at 70K. Reader interpretation: shared computation is an architectural fact, whereas useful transfer is an empirical property that must be checked for each direction and regime. homogeneous-transfer
5.3 Read upstream transfer as a change in capabilities, not one score
The cascaded experiments provide an intermediate explanation for the final MOTS result. With Mask R-CNN, alternating detection and instance-segmentation batches adds location supervision to mask learning. The tracking experiment uses a shared DLA-34 base with modified Faster R-CNN and gains detection AP and MOTA, even as identity switches increase. MOTS then combines localization, segmentation, and temporal association in one downstream evaluation. Its instance-pretrained and MOT-pretrained rows affect precision and identity errors differently, and the four-task row leads on AP and MOTSA without leading on every component metric. Reader interpretation: upstream tasks supply partly complementary capabilities, but the final comparison also changes initialization and data exposure. A reproduction should therefore inspect the error vector and use matched controls before assigning the gain to diversity or architectural sharing. instseg-transfermot-transfermots-transferprotocol-boundaries
5.4 Training and inference
During training
Lane/drivable models use DLA-34 and heads with four 3×3 convolution blocks followed by a 1×1 convolution. Three lane heads predict the attributes. Lane training uses weighted cross-entropy with foreground weight 10; experiments use 10K, 20K, or 70K images. homogeneous-transfer
Instance segmentation uses Mask R-CNN with ResNet-50, alternating detection and segmentation batches round-robin. Detection/MOT uses modified Faster R-CNN with a shared DLA-34 backbone. The final MOTS condition initializes from the jointly trained detection/tracking model and then jointly trains all four tasks; data and initialization differ from MOTS-only training. instseg-transfermot-transfermots-transfer
During inference
Dense heads predict maps at one-quarter spatial resolution and bilinearly upsample. Lane post-processing uses gradient-based nonmaximum suppression; evaluation additionally applies morphological thinning at each score threshold. Tracking predicts temporal object associations, but the main paper defers implementation details to supplementary material. homogeneous-transfermot-transfer
No learned future-world rollout or inference-time action-selection loop is established. GPS/IMU is proposed as imitation supervision with perplexity-based behavior comparison, whose protocol is referred to another paper. Perception metrics do not measure executed vehicle control. imitationmots-transfer
5.5 Implementation flow
- Collect and split videos
The collection contains 100K forty-second videos at 720p and 30 fps, from more than 50K rides. GPS/IMU preserves trajectories. Videos are divided into 70K/10K/20K training/validation/test sets; image tasks annotate the tenth-second frame. New York and the San Francisco Bay Area are prominent collection regions. collection
- Define distinct annotation targets
Image tags describe weather, scene, and time. Detection has ten categories plus occlusion/truncation attributes. Lane labels encode category, continuity, and direction. Directly drivable regions describe the current driving area and priority; alternative regions require changing lanes. These affordance labels depend on surrounding objects and context. annotation-semantics
- Reserve costly labels for smaller subsets
Semantic/instance annotation samples 10K clips, split 7K/1K/2K, with forty annotation classes and instance IDs for suitable classes. MOT uses 2,000 videos split 1,400/200/400, annotated at 5 fps. MOTS uses ninety videos split 60/10/20. These subsets must not be conflated with the 100K image-task split. segmentation-datatracking-data
- Compare different routes for supervision
Homogeneous experiments share dense prediction tasks. Cascaded experiments add detection labels to instance segmentation or tracking. Heterogeneous MOTS experiments first transfer upstream models, then combine detection, tracking, instance segmentation, and MOTS training. homogeneous-transferinstseg-transfermot-transfermots-transfer
6. Experiments & results
BDD100K organizes driving videos into annotation sets with different costs and output structures. Its experiments ask whether abundant simpler labels improve scarce, harder perception tasks. Detection supervision helps instance segmentation, tracking, and semantic segmentation, but transfer depends on the auxiliary task and available data. The combined MOTS result improves aggregate accuracy while retaining substantial errors; it establishes perception transfer, not autonomous driving competence.
This dataset paper has annotation illustrations, distribution plots, and task-transfer tables, but no dedicated network architecture figure. Figures 1 and 6 illustrate the input/label mechanism, while Section 5 supplies the baseline architecture description. The paper does not provide an executed driving-control experiment; Section 3.8 defers imitation-learning protocol details to another work. Neither a control-result visual nor a reconstructed architecture is substituted. task-overviewannotation-semanticshomogeneous-transfermot-transferimitation
6.1 Read the original evidence
Table 4. Equal training-set size does not remove domain dependence. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat each row as a training condition and each column as an evaluation domain. The left block varies city versus non-city imagery; the right block varies time of day. Every training subset contains 30K images, and Section 4.1 uses ResNet-50 Faster R-CNN with the COCO evaluation API. To assess a training choice, compare rows within one column: for example, compare Daytime-30K against Random-30K in Non-Daytime. To see domain sensitivity within one model, compare columns along a row. Keep the Val summary column distinct from either domain-specific column and retain the source's exact Non-Daytime label. domain-discrepancycollection
What it supports. The daytime-trained model scores 30.6 AP on Daytime but 23.6 on Non-Daytime. Random sampling gives 26.0 on Non-Daytime, compared with 23.6 for daytime-only training, while Daytime AP changes from 30.6 to 29.5. The collection therefore exposes a tradeoff hidden by a single aggregate score.
Where the evidence stops. The columns do not contain identical images, so a within-row gap mixes domain shift and evaluation-set difficulty. Equal image counts do not equalize object composition or ride diversity. The table supplies no uncertainty or per-domain evaluation sample counts.
Table 9. The best aggregate MOTS score does not minimize every error. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Decode the row abbreviations before comparing results: S is MOTS, I is instance segmentation, T is multiple-object tracking, and Det is detection. Section 5.3 describes the middle rows as fine-tuning MOTS from upstream instance or MOT models. The bottom row instead starts from the joint detection/tracking model and then trains all four tasks together. Read AP and MOTSA alongside MOTSP and the error counts: FN denotes false negatives, FP false positives, and IDSW identity switches. A higher aggregate score can coexist with worsening individual errors, and the best starting task can depend on which failure mode matters. mots-transferprotocol-boundariesimitation
What it supports. The combined condition reaches 23.3 AP and 41.4 MOTSA, versus 13.0 and 30.4 for MOTS alone. False negatives fall from 8352 to 5132, while false positives rise from 5116 to 6228. MOT initialization has fewer identity switches than the combined condition, 390 versus 472. Transfer improves different capabilities unevenly.
Where the evidence stops. The bottom row changes training data, task mix, and initialization. It therefore cannot by itself identify which component causes the gain. MOTSA is an offline perception metric; no closed-loop driving success or safety improvement is measured by this table.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Multiple object tracking and segmentation BDD100K MOTS; 12K training frames from 60 videos plus auxiliary detection, MOT, and instance labels; evaluation split not explicitly identified. | All four tasks: 23.3 AP, 41.4 MOTSA, 81.6 MOTSP; 5132/6228/472 errors. Instance AP; MOTSA; MOTSP; FN/FP/IDSW | MOTS-only: 13.0 AP, 30.4 MOTSA, 81.8 MOTSP; 8352/5116/566 errors. Aggregate accuracy rises; false positives rise and MOTSP falls slightly. The 11.0 MOTSA-point gain is a calculated difference, not a controlled estimate of diversity alone. mots-transfer |
| Lane marking and drivable-area joint learning BDD100K DLA-34; matched 10K or 70K training images; evaluation split unstated. | Joint 10K: 50.40 lane, 64.37 drivable; joint 70K: 54.24 lane, 72.21 drivable. Mean lane ODS-F at tolerance 10 pixels (%); mean drivable IoU (%) | Single-task 10K: 45.41 lane, 64.23 drivable; 70K: 54.48 lane, 71.37 drivable. Lane transfer helps at 10K but does not persist at 70K. No uncertainty is supplied. homogeneous-transfer |
| Instance segmentation with detection supervision BDD100K Mask R-CNN/ResNet-50; 7K instance images plus 70K detection images; evaluation split unstated. | Joint: 24.5; 45.4; 21.6. Mask AP; AP50; AP75 | Instance-only: 21.8; 40.5; 20.5. Additional box supervision improves segmentation without isolating additional data from task-sharing effects. instseg-transfer |
| Multiple object tracking with detection supervision BDD100K modified Faster R-CNN/DLA-34; 278K MOT training frames plus 70K detection images; evaluation split unstated. | Joint: 30.7; 56.7; 84.1; 9098. Detection AP; MOTA; MOTP; identity switches | MOT-only: 28.1; 55.0; 84.0; 8386. Detection and overall tracking accuracy improve while identity consistency worsens. mot-transfer |
| Semantic segmentation with auxiliary tasks BDD100K semantic model fine-tuned jointly with detection or lane/drivable supervision; evaluation split unstated. | With detection: 58.3; with lane/drivable: 53.3. Mean IoU | Semantic-only: 56.9. Road/sidewalk improve with lane/drivable labels, but aggregate segmentation declines. Similar output structures do not ensure beneficial transfer. semantic-transfer |
| Object detection across time-of-day domains ResNet-50 Faster R-CNN; 30K training images per condition; BDD100K domain evaluation and Val column, COCO API. | Daytime-trained: 30.6 on Daytime, 23.6 on Non-Daytime, 28.1 on Val. Detection AP | Random-30K: 29.5, 26.0, 28.3 on the same columns. Mixed-domain training improves Non-Daytime AP over daytime-only training. Non-Daytime is the table label; it should not be silently narrowed to night. domain-discrepancy |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 5. Lane transfer is strongest when labeled training data is scarce. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read one size block at a time, comparing Lane against Lane+Drive in the lane mean column and Drive against Lane+Drive in the drivable mean column. The lane subcolumns are direction, continuity, and category; their mean is separate from the directly/alternatively drivable IoU mean. The lane header fixes the matching tolerance at ten pixels. Section 5.1 additionally specifies thinning at each score threshold, so these values should not be compared with an evaluation that changes contour processing. After making within-block comparisons, move down the table to see how transfer changes as the same task pair receives more training images. homogeneous-transferprotocol-boundaries
What it supports. At 10K images, joint training raises lane mean ODS-F from 45.41 to 50.40, while drivable mean IoU barely changes, from 64.23 to 64.37. At 70K, lane mean is 54.24 jointly versus 54.48 alone. The visible benefit is conditional on task direction and data scale.
Where the evidence stops. The authors' explanation that similar output structures provide little new information is a hypothesis. This table does not isolate representation overlap, optimization effects, or training variance; it also does not explicitly identify validation versus test evaluation.
Table 8. Related spatial labels can improve road classes while harming the overall task. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin at the far-right mean IoU column, then inspect the category columns to understand what that average combines. The first row is semantic segmentation alone. The second adds detection training; the third adds lane and drivable-area training. Compare road and sidewalk with rider and bicycle rather than reading boldface as a universal winner. Section 5.2 describes fine-tuning a base semantic model with these auxiliary tasks. The source's annotation inventory contains forty classes, but this evaluation table displays nineteen categories; the main text does not provide a complete mapping, so the reported average should be read in the table's own evaluation scope. semantic-transfersegmentation-dataprotocol-boundaries
What it supports. Detection raises mean IoU from 56.9 to 58.3. Lane/drivable supervision lowers it to 53.3 despite improving road from 94.3 to 94.8 and sidewalk from 63.0 to 65.8. Rider falls from 22.1 to 7.1, and bicycle from 49.9 to 27.5. Local improvements can mask a worse overall model.
Where the evidence stops. This comparison demonstrates an outcome for the reported training choices, not an inevitable conflict between these tasks. Loss balancing, category mapping, training exposure, and variance are not sufficiently controlled here to identify a unique cause of the regressions.
7. Analysis & limitations
7.1 What the evidence leaves open
Geographic variety is concentrated in populated US regions, and detection categories have a long tail. Diversity within this collection does not establish worldwide coverage or balanced recognition quality. collectionannotation-semantics
Tables 5–9 provide point estimates without uncertainty and do not explicitly name their evaluation split. Upstream-data comparisons change supervision amount, video diversity, and sometimes initialization together; causal attribution to any one factor remains unresolved. protocol-boundariesmots-transfer
The annotation inventory has forty semantic classes, while Table 8 reports nineteen evaluated categories. The main text gives no complete mapping between these sets. Driver/ride separation across splits is also unspecified, so leakage protection cannot be assumed. collectionsegmentation-datasemantic-transfer
7.2 Questions for discussion
- Does diversity across videos help MOTS more than additional annotated frames from the same rides? [tracking-data, mots-transfer]
- Which auxiliary losses cause the semantic class regressions, and do they persist under matched training exposure? [semantic-transfer]
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs original task subsets, category mappings, initialization, task sampling and loss weights, evaluation split, and matching metrics. Lane evaluation specifies tolerances 1, 2, and 10 pixels and thinning, but the supplied paper lacks a complete optimizer/schedule, frozen-module policy, hardware budget, and tracking implementation. homogeneous-transferprotocol-boundariesmot-transfer
The title-page footnote gives a dataset URL, but the supplied paper does not state dataset license terms. A URL alone establishes neither current access nor an inspected implementation. availability
Proposed checks: equalize training exposure when testing upstream transfer, and repeat the homogeneous comparison across data sizes with fixed evaluation and multiple seeds. These test whether benefits survive controls missing from the reported comparisons. homogeneous-transfermots-transfer
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: Disentangle upstream data from initialization in MOTS
Reader-proposed experiment, not performed: fix the MOTS training subset and a declared held-out evaluation split, then cross MOTS-only versus four-task training with common initialization versus the paper's detection/MOT initialization. Include separate controls matching total optimizer updates and matching MOTS batch exposure; use MOTS-only repetition to control for extra training. Use identical architecture, augmentations, and evaluation code wherever the missing implementation details can be resolved. Repeat across seeds and report AP, MOTSA, MOTSP, FN, FP, and IDSW. An advantage that disappears after matched exposure would weaken the claim that auxiliary supervision itself explains the gain; a persistent reduction in FN with increased FP would reproduce the reported tradeoff rather than universal improvement. tracking-datamots-transferprotocol-boundaries
Check 2: Test whether the lane benefit really depends on label scarcity
Reader-proposed experiment, not performed: choose nested 10K, 20K, and 70K training subsets and train lane-only, drive-only, and joint DLA-34 models with matched image exposure, schedules, and several random seeds. Hold preprocessing, lane foreground weighting, and nonmaximum suppression constant. Reproduce Table 5 at ten-pixel tolerance, then report the paper's one- and two-pixel tolerances with identical thinning. Compare per-attribute lane scores and direct/alternative IoU, not just means. A consistent lane gain only at 10K would support the reported scarcity-dependent pattern; a gain that vanishes under tighter tolerance would show that the improvement does not establish equally precise boundaries. homogeneous-transferprotocol-boundaries
8.3 Reading coverage
Visual audit: All four text chunks were read completely. All ten physical PDF pages were rendered and visually inspected, including title/affiliations and version watermark, annotation figures, distribution plots, every quantitative table, method and evaluation text, and references. Six final original crops were individually viewed. Figures 1 and 6 were checked against their captions and label definitions; table row/column alignment and numerical tradeoffs were checked on the full pages and crops. The paper contains no dedicated network architecture diagram; the method account follows Section 5. No supplementary pages, linked resources, dataset files, or executed experiments are included in this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10. 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 Works
- 3. BDD100K; 3.1 Image Tagging; 3.2 Object Detection; 3.3 Lane Marking; 3.4 Drivable Area; 3.5 Semantic Instance Segmentation; 3.6 Multiple Object Tracking; 3.7 Multiple Object Tracking and Segmentation; 3.8 Imitation Learning
- 4. Diversity; 4.1 Object Detection; 4.2 Semantic Segmentation
- 5. Multitask Learning; 5.1 Homogeneous Multitask Learning; 5.2 Cascaded Multitask Learning; 5.3 Heterogeneous Multitask Learning
- 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.
- Identity/version: the title and all eight authors match the catalog. The inspected artifact identifies itself as the CVPR 2020 CVF Open Access version, identical to the accepted version except for its watermark. No numbered revision or revision history is supplied. Printed pages 2636–2645 differ from the catalog BibTeX range 2633–2642; citations here use physical PDF pages. No other edition was inspected.
- Text extraction does not reconstruct figure images; this limitation was addressed by inspecting all ten PDF pages and every final crop.
- Separate supplemental material availability has not been fully verified. No supplement was supplied; detailed annotation distributions, trajectory visualizations, full lane evaluations, and tracking implementation details deferred there were not read.
- Code, dataset files, linked resources, and cited papers were not inspected. No experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
identityPDF p. 1, title/authors, affiliation line and CVF watermark; pp. 1 and 10, printed footers
Exact title and eight authors match. Affiliations are UC Berkeley, Cornell University, UC San Diego, and Element, Inc.; starred work was done at UC Berkeley. CVF identifies its CVPR 2020 Open Access version; observed pagination is 2636–2645.
Go to primary source ↓task-overviewPDF pp. 1–2, Abstract, Section 1, Figure 1 and caption
BDD100K contributes a driving-video dataset and heterogeneous multitask benchmark. The introduction enumerates ten task directions; the overview shows annotation types and temporal segmentation examples.
Go to primary source ↓collectionPDF p. 3, Section 3 and Figure 2
100K videos, forty seconds, 720p/30 fps, more than 50K rides, GPS/IMU, 70K/10K/20K split, tenth-second image annotation. Collection includes New York and the Bay Area; ride-disjoint splitting is not specified.
Go to primary source ↓annotation-semanticsPDF p. 3, Sections 3.1–3.2 and Figures 3–4; p. 4, Sections 3.3–3.4 and Figures 5–6
Image tags, ten box categories and visibility attributes; long-tailed counts. Lane categories/continuity/direction differ from directly and alternatively drivable region semantics. Figure 6 identifies red direct and blue alternative regions.
Go to primary source ↓segmentation-dataPDF p. 4, Section 3.5; p. 5, continuation of Section 3.5
Pixel labels cover forty annotation classes on 10K sampled clips, with instance IDs for appropriate classes and 7K/1K/2K image splits.
Go to primary source ↓tracking-dataPDF p. 5, Sections 3.6–3.7 and Tables 2–3; p. 8, Sections 5.2–5.3
MOT: 2,000 videos, 5 fps, 1,400/200/400 split; training has 278K frames. MOTS: ninety videos, 60/10/20 split; training has 12K frames. Table 2 is training plus validation, not the full collection.
Go to primary source ↓tracking-scalePDF p. 5, Section 3.6, Figure 7 and caption
Box size is sqrt(wh), with w and h the box width and height. The cumulative distributions describe size, within-track size variation, and track length.
Go to primary source ↓imitationPDF p. 6, Section 3.8
GPS/IMU is proposed for demonstration supervision and perplexity-based comparison; evaluation details are referred to Xu et al. [32] and trajectories to supplementary material. No executed-control result is supplied here.
Go to primary source ↓domain-discrepancyPDF p. 6, Table 4, Daytime-30K and Random-30K rows; Sections 4.1–4.2 and Figure 9
Equal 30K training conditions use ResNet-50 Faster R-CNN and COCO API. Daytime-trained AP is 30.6/23.6/28.1 versus random 29.5/26.0/28.3 on Daytime/Non-Daytime/Val. Figure 9 qualitatively compares DRN models across datasets.
Go to primary source ↓homogeneous-transferPDF p. 7, Section 5.1 and Table 5, 10K/20K/70K blocks
DLA-34, convolutional heads, lane foreground weight 10, interpolation/NMS, ODS-F tolerances and thinning are specified. Table 5 contains single/joint lane and drivable values; the authors hypothesize limited new information from similar output structures.
Go to primary source ↓instseg-transferPDF p. 7, Section 5.2, Object detection and instance segmentation; Table 6
Mask R-CNN/ResNet-50 alternates 7K instance and 70K detection data by batches. AP/AP50/AP75 change from 21.8/40.5/20.5 to 24.5/45.4/21.6.
Go to primary source ↓mot-transferPDF p. 7, Table 7; p. 8, Section 5.2, MOT and object detection
Modified Faster R-CNN shares DLA-34 for MOT/detection; implementation is deferred to supplement. AP/MOTA/MOTP/IDS change from 28.1/55.0/84.0/8386 to 30.7/56.7/84.1/9098.
Go to primary source ↓semantic-transferPDF p. 8, Table 8, all rows and mean IoU, road, sidewalk, rider, bicycle columns; Section 5.2
Mean IoU is 56.9 alone, 58.3 with detection, 53.3 with lane/drivable. Road/sidewalk change from 94.3/63.0 to 94.8/65.8 with lane/drivable, while rider/bicycle fall from 22.1/49.9 to 7.1/27.5. Nineteen category columns are shown.
Go to primary source ↓mots-transferPDF p. 8, Section 5.3 and Table 9, all rows
MOTS-only, instance-pretrained, MOT-pretrained, and four-task conditions report AP/MOTSA 13.0/30.4, 18.7/33.7, 19.7/40.3, 23.3/41.4. Four-task initialization uses the joint detection/MOT model. Errors improve nonuniformly; MOT initialization has 390 IDSW versus 472 for all tasks.
Go to primary source ↓protocol-boundariesPDF pp. 7–8, Sections 5.1–5.3 and Tables 5–9
Training subsets and selected model details are described, but table evaluation splits, uncertainty, full schedules, hardware budget, and a complete freezing policy are not explicitly supplied. Auxiliary-data conditions do not isolate equal supervision budgets.
Go to primary source ↓availabilityPDF p. 1, dataset footnote; pp. 3–8, supplement references in Sections 3.1, 3.3–3.8, 5.1–5.2
The paper points to a dataset URL without specifying license terms; annotation distributions, trajectory plots, full lane evaluations, and tracking implementation are deferred outside this PDF.
Go to primary source ↓8.5 Primary sources
BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning ↗
PDF · 6,872 extracted words
Source fingerprint
5123a1ac802f639a87975199ed9972463b5faa43fcbade3ef58de9a26f06531e