Semi-Supervised Deep Image Stitching for Moving Elongated Objects

Abstract

Image-stitching methods for moving elongated objects require high stitching quality, efficient inference, and robustness to interference from regions outside the target object. Existing methods still have difficulty satisfying these requirements simultaneously. This paper proposes a semi-supervised deep image-stitching method for moving elongated objects. The proposed framework consists of two stages: semi-supervised registration and unsupervised reconstruction. In the semi-supervised registration stage, a semi-supervised optical-flow estimation network is used to predict the bidirectional optical flow between the input images. An object-centric spatial transformation module is then introduced to remove regions outside the moving object and warp the inputs onto a unified plane. In the unsupervised reconstruction stage, a multi-scale fusion model is used to improve the quality of the reconstructed stitched image. Correspondingly, we design a reconstruction objective function based on multi-scale feature representations. To address the lack of available datasets for this task, we construct two datasets: MEOIS-D, a synthetic dataset for generalized evaluation, and Container-D, a real-world scene-specific dataset. Extensive comparative experiments and ablation studies demonstrate the effectiveness of the proposed method.