Research map / Detection and segmentation / Weakly supervised detection and segmentation
A New Multi-Source Light Detection Benchmark and Semi-Supervised Focal Light Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces YDLD, a 3.5k-image, 116k-box multi-source light detection dataset from YouTube driving videos, and a semi-supervised focal light detector (SS-FLD) using a lightness focal loss and a lightness spatial attention prior.
Problem
Multi-source light detection (vehicle lights, traffic signal lights, streetlights) in driving scenes is understudied, mainly due to the lack of public multi-class benchmark datasets; existing detectors give low mAP on tiny, similar-looking lights.
Approach
A lightness focal loss that adds a false-positive term 1/(q+eps) and a geometry-based lightness spatial attention prior (learned from GT box density relative to vanishing points), embedded into a semi-supervised object detection (mean-teacher) framework.
Key results
- SS-FLD with RFLA w/t PAA reaches 26.0 mAP on YDLD, the best among ~20 detectors (DINO baseline 22.6).
- Lightness focal loss plus SSOD improves RFLA from 21.6 to 26.0 mAP and PAA from 21.6 to 25.6 mAP.
- Pre-training DINO on YDLD gives transfer gains on BDD100k that match or exceed COCO pre-training (38.3 mAP) despite fewer overlapping classes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pretrain DINO on YDLD before fine-tuning on BDD100k | DINO · BDD100k | mAP | Helped |
| Add LF loss to RFLA (ablation) | RFLA · YDLD | mAP: 21.6 -> 24.3 | Helped |
| Add SSOD to RFLA with LF loss (ablation) | RFLA · YDLD | mAP: 24.3 -> 26.0 | Helped |
| Add LF loss to PAA (ablation) | PAA · YDLD | mAP: 21.6 -> 24.5 | Helped |
| Add SSOD to PAA with LF loss (ablation) | PAA · YDLD | mAP: 24.5 -> 25.6 | Helped |
| Add the FP focal loss term (ablation) | Original focal loss in M1 · YDLD | mAP: 21.6 -> 24.4 | Helped |
| Add the lightness spatial attention prior (ablation) | Original focal loss in M1 · YDLD | mAP: 21.6 -> 24.5 | Helped |
| Add both the FP focal loss and lightness spatial attention prior (ablation) | Original focal loss in M1 · YDLD | mAP: 21.6 -> 25.6 | Helped |
| Use SS-FLD with RFLA instead of the RFLA baseline | RFLA · YDLD test | mAP: 21.6 -> 26.0 | Helped |
| Use SS-FLD with DINO instead of DINO | DINO · YDLD test | mAP: 22.6 -> 25.6 | Helped |
Surprising findings
- Pre-training on YDLD (3.5k images) transfers to BDD100k as well as COCO pre-training does, despite having the fewest overlapping classes with BDD100k.
- SoftTeacher scores only 8.4 mAP on YDLD, below several non-semi-supervised detectors.
Limitations
- Dataset has only three light classes; indoor lights, neon signs, and sunlight are excluded.
- Very tiny objects keep mAP low (APvt around 12%), so localization remains hard.
- Hyperparameters (eta, alpha, gamma) tuned on the benchmark; the lightness prior depends on fixed-camera, flat-road geometry.
Experimental setup
Datasets: YDLD BDD100k KITTI SODA10M ImageNet MS-COCO
Benchmarks: YDLD BDD100k
Metrics: box AP (mAP at IoU 0.5:0.95) APvt APt APs AP50 AP75
Baselines: Libra R-CNN Faster R-CNN Cascade R-CNN YOLOF RetinaNet ATSS YOLOX PAA Deformable DETR DINO SoftTeacher MeanTeacher
Builds on: focal loss DINO PAA RFLA ConsistentTeacher
Closest papers by meaning
- You Only Look Around: Learning Illumination-Invariant Feature for Low-light Object Detection (NeurIPS 2024 · similarity 0.85)
- Light the Night: A Multi-Condition Diffusion Framework for Unpaired Low-Light Enhancement in Autonomous Driving (CVPR 2024 · similarity 0.84)
- Trash to Treasure: Low-Light Object Detection via Decomposition-and-Aggregation (AAAI 2024 · similarity 0.84)
- Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · similarity 0.84)
- FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network (NeurIPS 2025 · similarity 0.83)
- Towards RAW Object Detection in Diverse Conditions (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.