Efficient Local Correlation Volume for Unsupervised Optical Flow Estimation on Small Moving Objects in Large Satellite Images

Sarra Khairi, Etienne Meunier, Renaud Fraisse, Patrick Bouthemy; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 440-448

With the advent of deep learning methods performance and efficiency of optical flow estimation has significantly increased especially for supervised models. However they do not generalize well to more specific data involving small moving objects in large images such as high-resolution aerial or satellite sequences. In addition annotation and realistic simulation are difficult for these contents which calls for unsupervised alternatives. Yet the latter are still less accurate than their supervised counterparts. In this paper we introduce an unsupervised local optical flow estimation method adapted to small moving objects in large-size images by involving no downsampling of the feature maps. We adopt a local correlation search and implement it in an original way with a per-shift computation which minimizes memory consumption and speed up inference computation for large-scale images. We also design a loss function combining similarity smoothness and sparsity constraints. We demonstrate the performance of our SMOFlow method on real stabilized aerial videos fully representative of future satellite conditions. SMOFlow favorably compares to other methods. Our SMOFlow method is able to accurately capture the motion of small objects in large images while efficiently reducing memory consumption.