Sending local data to cloud servers is vulnerable to user privacy, and its long update latency. Meanwhile, the state-of- the-art stereo matching method is still computation demanding, fine-tuning the whole model on-device is not a practicable solution because of the limited power budget and computation ability on edge devices. In this study, we propose a two-stage online stereo matching refinement system, using an additional light-weight network to learn the domain gap between local data and cloud training data. We define a load-gain ratio to evaluate computer efficiency. This refinement system has a much better load-gain ratio than fine-tune. (0.2 v.s. 35.7 operation overhead/accuracy gain) Nevertheless, we only disburse 0.2% of additional parameters and 0.7% additional computation as set by inference the stereo matching model. Thus, it would be a suitable choice for an online training scenario. With re-scheduling the training pipeline, we use a patch-based layer fusion technique and reduce the off-chip memory bandwidth by 97%.