| Title |
Adaptive Class-aware Transfer Learning for Semantic Segmentation in Off-road Autonomous Driving |
| Authors |
류제호(Je-ho Ryu) ; 김용휘(Yong-hwi Kim) ; 이승주(SeungJoo Lee) ; 임태윤(Tae-Yoon Lim) ; 손호정(Ho-Jung Sohn) ; 조용진(Yong-Jin Jo) ; 조지혁(Jihyuk Cho) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.4.91 |
| Keywords |
Semantic segmentation; Transfer-learning; Off-road environment; Class imbalance; Domain adaptation |
| Abstract |
Semantic segmentation in off-road environments faces limitations in securing perception performance for autonomous driving compared to urban environments, primarily due to diverse unstructured objects and unbalanced data distribution among classes. To overcome this, transfer learning based on large-scale dataset is essential, however, the scarcity of off-road dataset induces feature distortion and overfitting in learning models, while data bias degrades the detection performance for minority objects. Accordingly, this paper proposes Adaptive Class-aware Transfer for Segmentation (ACT-Seg), a transfer learning methodology for off-road environments that optimizes learning parameters and loss functions. ACT-Seg is an adaptive transfer-learning framework that selectively switches the trainable scope according to target data scale. Specifically, a validation-selected threshold τ determines whether to freeze the backbone and train only the segmentation head or to perform full fine-tuning. In addition, ACT-Seg employs a class-imbalance mitigation loss configuration that combines focal loss with effective-number-based class weights. To validate the effectiveness of the proposed method, transfer learning experiments were conducted using a semantic segmentation model pre-trained on the GOOSE dataset (source domain) and adapted to the RELLIS-3D dataset (target domain). Experimental results demonstrate that the proposed method significantly improved the mean Intersection over Union (mIoU) on the target domain compared to existing methods and enhanced detection performance for minority classes. |