VesselDataset: RGB-Infrared Aligned Vessel Segmentation with Iterative Noise Filtering Strategy

A high-quality multi-modal vessel dataset is urgently needed to advance maritime computer vision research, yet current resources remain limited. In this work, we present a novel vessel dataset featuring aligned RGB and infrared images captured from UAV perspectives, with both instance-level and pixel-level annotations for vessel detection and segmentation tasks. To efficiently construct this large-scale dataset, we employ pseudo-labels for semantic segmentation, which inevitably introduces annotation noise. To address this challenge, we propose an iterative noise filtering strategy that leverages a novel moving-average filtering criterion and a drop head-and-tail operation. This approach maintains a moving average of model predictions across epochs, stabilizing the noise filtering process and effectively distinguishing between noisy and clean samples. Unlike traditional methods that only remove high-loss samples, our strategy also discards certain low-loss samples, which are often dominated by background pixels in segmentation tasks. Furthermore, re-annotating discarded samples with improved pseudo-labels further enhances model performance. Extensive experiments on our curated noisy semantic segmentation datasets demonstrate the effectiveness of the proposed iterative noise filtering approach in improving segmentation quality.