TWDetector: A Diffusion-Based Detection for Two-Wheelers in Low-Light Environments
Low illumination can obscure the boundaries and fine visual cues of small two-wheelers. TWDetector combines Dynamic Tanh-based Swin feature processing and bidirectional multi-scale fusion with diffusion-based box refinement for iterative two-wheeler detection. On the BDD100K-derived Data-TWC validation set, the full configuration achieves 25.540% AP and 11.231% APs, compared with 26.056% AP and 9.417% APs for the diffusion-only configuration. These results describe the trade-off between overall and small-object AP on the selected validation split under the respective training recipes. On the 343-image CODA-derived external subset, TWDetector achieves 7.391% AP and 9.174% APs. A descriptive analysis of the 186 images labeled night yields 6.753% AP and 3.805% APs, both below the diffusion-only model, limiting claims of improved nighttime generalization. The results characterize the accuracy trade-off of the evaluated configuration and identify cross-dataset robustness and inference efficiency as priorities for further development.
Authors
- Siyu Teng (ORCID: https://orcid.org/0000-0002-6860-9547)
- Ying Yang
- Haijun Li
- Guang Yang
- Yang Liu
- Jiasong Zhu
- Long Chen
Institutions
- Shenzhen University (CN)
- Chinese Academy of Sciences (CN)
- Guangzhou Urban Planning Survey & Design Institute (CN)
Publication Details
- Journal
- Vehicles
- Published
- 2026-11-17
- DOI
- https://doi.org/10.3390/vehicles8110303
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China
- National Key Research and Development Program of China