Removing Cost Volumes from Optical Flow Estimators

https://visinf.github.io/recover Overlayed input images SEA-RAFT [52] ReCoVEr-CX (ours) PWC-Net [47, 48] ReCoVEr-MN (ours) inference: 169ms 8.22GB inference: 144ms 1.24GB inference: 43ms 1.25GB inference: 50ms 0.49GB Overlayed input images SEA-RAFT [52] inference: 169ms 8.22GB ages SEA-RAFT [52] ReCoVEr-CX (our inference: 169ms 8.22GB inference A-RAFT [52] ReCoVEr-CX (ours) PWCinference: 169ms 8.22GB inference: 144ms 1.24GB ReCoVEr-CX (ours) PWC-Net [47, 48] inference: 144ms 1.24GB inference: 43ms 1.25GB urs) PWC-Net [47, 48] ReCoVEr-MN (our ce: 144ms 1.24GB inference: 43ms 1.25GB inferenc Overlayed input images SEA-RAFT [52] inference: 169ms 8.22GB images SEA-RAFT [52] ReCoVEr-CX (ou inference: 169ms 8.22GB inferenc EA-RAFT [52] ReCoVEr-CX (ours) PWC inference: 169ms 8.22GB inference: 144ms 1.24GB ReCoVEr-CX (ours) PWC-Net [47, 48] inference: 144ms 1.24GB inference: 43ms 1.25GB ours) PWC-Net [47, 48] ReCoVEr-MN (ou nce: 144ms 1.24GB inference: 43ms 1.25GB inferen Figure 1. ReCoVEr. We propose a method to remove cost volumes from optical flow estimators during training, and thereby, we are able to create fast and accurate optical flow estimators with a significantly reduced memory footprint. Our most accurate model, ReCoVEr-CX, reaches state-of-the-art accuracy while being more efficient w.r.t. inference and memory than SEA-RAFT [52]. Our most efficient model, ReCoVEr-MN, predicts sharper motion boundaries compared to the popular PWC-Net [47, 48], while having comparable efficiency.