Beyond RGB: Channel Fusion for Improved Cross-Domain Car Detection

Domain adaptation in cross-domain vision has significantly improved detection performance under adverse weather and visibility conditions. However, this enhancement has come at the cost of substantial complexity in the training routine and an increase in computational cost. In this article; a crossdomain car detection framework is proposed that outperforms state-of-the-art domain-adaptive detectors while maintaining a remarkably lower model complexity and computational overhead. The proposed model architecture is based on YOLOv8 with the integration of CBAM (Convolutional Block Attention Module). In addition, a novel channel combination is proposed where the ‘Y’(Luminance) channel from the YCbCr color space is fused with the ‘G’(Green) and ‘B’(Blue) channels from the RGB color space producing YGB. The BDD100K data set is used for model training and validation to adapt the model to a driving context where it can be deployed in autonomous cars. The proposed framework outperforms state-of-the-art models in cross-domain car detection by at least 3.95 % in precision, and at least 11.81 % in F1 score. When compared to other equally performing state-of-the-art domain adaptation techniques, it showed drastic decrease in the number of trainable parameters (up to 95.2%).