We propose a noisy-label resilient model-agnostic training framework named Active Noise Cancellation (ANC) for semantic segmentation in spatial computing. In the presence of noisy labels, which arise due to measurement errors, crowdsourcing, insufficient expertise and so on, deep learning tends to have poor generalization on test data. In spatial computing, noisy labels largely come from crowdsourcing and measurement error. The ANC framework is a training paradigm to improve label quality during batch update, detect the unreliable pixel labels and filter them during training. We demonstrate the effectiveness of our proposed framework on two satellite image datasets for building footprint detection. As a result, our method produces better intersection over union (IoU), precision, recall and F1 score when training with noisy masks.