Viewpoint-Aware Sampling for Effective Online Domain Incremental Learning
Mohammad Wasil, Tobias Glasmachers, Sebastian Houben · IFAC-PapersOnLine · 2025
We investigate the problem of online domain continual learning for image classification. Within an extended series of tasks, continual learning encounters the issue of catastrophic forgetting. To mitigate this challenge, one may employ a memory-replay strategy, a technique involving the re-visitation of stored samples in a buffer when new tasks are introduced. However, the memory budget available to autonomous agents, such as robots, is typically limited, making the selection of representative examples crucial. An effective strategy to ensure representativeness is to select diverse examples. To this end, we propose a novel on-the-fly sampling policy, called Viewpoint-Aware Sampling (VAS), which maintains diversity in the memory buffer by selecting examples from different visual perspectives. We empirically evaluate the effectiveness of VAS across the OpenLORIS-Object and the CORe50-NI benchmark and find that it consistently outperforms state-of-the-art methods in terms of average accuracy, backward transfer, and forward transfer, while requiring fewer computational resources.