Previous camera self-calibration methods have exhibited certain notable shortcomings. On the one hand, they either exclusively emphasized scene cues or solely focused on vehicle-related cues, resulting in a lack of adaptability to diverse scenarios and a limited number of effective features. Furthermore, these methods either solely utilized geometric features within traffic scenes or exclusively extracted semantic information, failing to comprehensively consider both aspects. This limited the comprehensive feature extraction from scenes, ultimately leading to a decrease in calibration accuracy. Additionally, conventional vanishing point-based self-calibration methods often required the design of additional edge-background models and manual parameter tuning, thereby increasing operational complexity and the potential for errors. Given these observed limitations, and in order to address these challenges, we propose an innovative roadside camera self-calibration model based on the Transformer architecture. This model possesses a unique capability to simultaneously learn scene features and vehicle features within traffic scenarios while considering both geometric and semantic information. Through this approach, our model can overcome the constraints of prior methods, enhancing calibration accuracy and robustness while reducing operational complexity and the potential for errors. Our method outperforms existing approaches on both real-world dataset scenarios and publicly available datasets, demonstrating the effectiveness of our approach.
Previous camera self-calibration methods have exhibited certain notable shortcomings. On one hand, they either exclusively emphasized scene cues or solely focused on vehicle-related cues, resulting in a lack of adaptability to diverse scenarios and a limited number of effective features. Furthermore, these methods either solely utilized geometric features within traffic scenes or exclusively extracted semantic information, failing to comprehensively consider both aspects. This limited the comprehensive feature extraction from scenes, ultimately leading to a decrease in calibration accuracy. Additionally, conventional vanishing point-based self-calibration methods often required the design of additional edge-background models and manual parameter tuning, thereby increasing operational complexity and the potential for errors. Given these observed limitations, and in order to address these challenges, we propose an innovative roadside camera self-calibration model based on the Transformer architecture. This model possesses a unique capability to simultaneously learn scene features and vehicle features within traffic scenarios while considering both geometric and semantic information. Through this approach, our model can overcome the constraints of prior methods, enhancing calibration accuracy and robustness while reducing operational complexity and the potential for errors. Our method outperforms existing approaches on both real-world dataset scenarios and publicly available datasets, demonstrating the effectiveness of our approach.
Unsupervised visual anomaly detection supports industrial quality inspection using defect-free reference images. Nearest-neighbor detectors can nevertheless assign high anomaly scores to normal regions when illumination or other acquisition conditions change. We propose Transformation-Response-Aware Correction for Anomaly Detection (TRAC-AD), a residual-scoring method built on a frozen DINOv2 backbone. Spatially paired feature differences between normal source images and their photometrically transformed views define a low-dimensional response subspace. During inference, TRAC-AD retrieves normal references in the original fused feature space, removes the response-subspace component of the deepest-layer matching residual, and combines its norm with unchanged scores from the other layers. On AeBAD-S, RobustAD-derived PCB and MetalParts settings, and MVTec AD 2 TESTpub, the mean gains in normalized area under the per-region overlap curve up to a 5% false-positive rate (PRO@0.05) are 3.20, 1.14, and 0.49 percentage points over the matched base pipeline, respectively. Original-resolution TESTpub evaluation yields a mean AucPro@0.05 gain of 0.90 percentage points. These results support using paired photometric responses to improve localization at low false-positive rates under acquisition changes.