FlowFormer++: Masked Cost Volume Autoencoding for Pretraining Optical Flow Estimation

Benchmark Model Rank Results
optical-flow-estimation-on-kitti-2015FlowFormer++ (tile) (extra data)–Fl-all: 4.52
optical-flow-estimation-on-kitti-2015-trainFlowFormer++ (C+T, zero-shot, tile) (extra data)–F1-all: 14.13EPE: 3.93
optical-flow-estimation-on-sintel-cleanFlowFormer++ (extra data)–Average End-Point Error: 1.07
optical-flow-estimation-on-sintel-finalFlowFormer++ (extra data)–Average End-Point Error: 1.94