Class-incremental visual scene understanding for multi-stage construction via sequential knowledge distillation and exemplar replay

Abstract

While visual scene understanding in construction must adapt to evolving environments where new-class objects consistently emerge throughout the project lifecycle, this critical challenge remains largely underexplored. This paper redefines class-incremental learning in the context of multi-stage construction scenarios and presents a pipeline that enables models to learn visual knowledge of new classes while retaining knowledge of previously learned ones. A multi-stage image dataset labeled only with new classes was assembled to reflect realistic, temporally evolving construction scenarios. The proposed pipeline was evaluated against baseline methods, including full retraining and fine-tuning, as well as an upper-bound model trained with access to both old and new class labels. The pipeline outperformed the baselines in object detection tasks, across all evaluation metrics and test stages. It even surpassed the upper-bound in several cases, despite its limited access to old-class labels. These findings highlight the potential of class-incremental learning for long-duration, temporally evolving environments.