DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework

Qualitative results at 1K, 2K, 4K, and 8K resolutions. (b) Results on the proposed Kubric-NK evaluation. Figure 1. (a) Optical flow results on multiple resolution images from the KITTI 2015 [36], Spring [35], DAVIS [40] and our proposed Kubric-NK datasets. Methods like FlowFormer++ [48] adopt tiling to handle high-resolution inputs, which causes square-shaped artifacts and loss of global context. Static approaches, such as SEA-RAFT [60], do not generalize well to large changes in resolution either. Our proposed DPFlow employs a flexible dual-pyramid design to adapt to larger inputs. (b) Our proposed Kubric-NK evaluation can quantify the generalization capabilities to resolutions up to 8K. At 8K, DPFlow outperforms RAPIDFlow [38] by 30% and FlowFormer++ by 83%.