Patterns of vehicle lights: Addressing complexities of camera-based vehicle light datasets and metrics
DOI: 10.1016/j.patrec.2024.01.003
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This paper addresses the complexities of representing and annotating vehicle lights in camera-based datasets for autonomous driving. The authors argue that accurate vehicle light detection is critical for three specific downstream tasks: nighttime vehicle detection (where lights reveal vehicle presence when bodies are obscured), 3D vehicle orientation estimation (leveraging the geometric symmetry of light pairs), and dynamic trajectory prediction (interpreting signals like braking or turning). The study highlights that different tasks require different data representations, creating a challenge for dataset curation and model training. To resolve these challenges, the authors introduce the LISA Vehicle Lights Dataset, an extended annotation of the ApolloCar3D dataset. The paper first reviews four primary methods for representing vehicle lights: bounding boxes, center points, corner keypoints, and segmentation masks. It analyzes the trade-offs of each, noting that bounding boxes are robust but imprecise, while masks are precise but lack geometric structure useful for pose estimation. The authors advocate for corner-based representations as a balance between precision and geometric utility. The LISA dataset was created by filtering ApolloCar3D to include only visible lights and annotating them with corner coordinates relative to the light’s center. The authors developed two cropping strategies: a "Vehicle-Only Context Approach," which crops the image to the vehicle and then centers the light, and a "Vehicle with Scene Context Approach," which centers the light within the full traffic scene. The final dataset contains 89,568 augmented examples of cropped vehicle lights with normalized corner offsets, alongside an additional 2,606 examples converted from segmentation labels in other datasets. This provides a large-scale resource specifically designed for corner regression tasks. Furthermore, the authors developed and evaluated Light Visibility Models to determine whether a specific light is visible in a given image, a necessary step for associating lights with vehicle instances in cascaded detection systems. Using binary cross-entropy loss, these visibility networks achieved over 90% accuracy across all light types. The paper concludes that standardized, task-specific annotations are essential for improving the reliability of autonomous driving systems, particularly in low-light conditions or when inferring vehicle intent. The LISA dataset and associated models are made publicly available to facilitate further research in this domain.
Key finding
The introduction of the LISA Vehicle Lights Dataset and associated Light Visibility Model provides a robust resource for training vehicle light detection models, achieving over 90% accuracy in predicting light visibility.
Methodology
dataset
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-28 |
| archive | success | canonical_url | — | — | 11 | 2026-06-06 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-04 |
| chunk | success | chunk | — | — | 1 | 2026-06-04 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-04 |
| enrich | success | — | — | — | 1 | 2026-05-28 |
| promote | success | — | — | — | 1 | 2026-06-04 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.