An Anisotropic Selection Scheme for Variational Optical Flow Methods with Order-Adaptive Regularisation

Mehl, Lukas; Beschle, Cedric Aaron; Barth, Andrea; Bruhn, Andrés

doi:10.1007/978-3-030-75549-2_12

Cited by 6 publications

(3 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The newly trained PWC-Net and IRR are substantially better than the respective, published models, with more than 2 percent reduction in average outlier percentage (Fl-all) on the KITTI 2015 benchmark. Both are also more accurate than some more recent models [29,52,53,60].…”

Section: Benchmark Resultsmentioning

confidence: 78%

“…A recent, notable architecture, RAFT [45], built a full cost volume and performs recurrent refinements at a single resolution. RAFT achieved a significant improvement over previous models on Sintel and KITTI benchmarks, and became a starting point for numerous new variants [19,26,29,36,48,51,52,58]. To reduce the memory cost of the all-pairs cost volume, Flow1D used 1D selfattention with 1D search, with minimal performance drop while enabling application to 4K video inputs [52].…”

Section: Previous Workmentioning

confidence: 99%

“…The widely-acclaimed RAFT model [45] reduced error rates on common benchmarks by up to 30% versus state-of-the-art baselines, outperforming PWC-Net by a wide margin. RAFT quickly became the predominant framework for optical flow [19,26,29,36,48,51,52,58] and related tasks [23,46]. The success of RAFT has been attributed primarily to its novel network design, including its multiscale all-pairs cost volume, its recurrent update operator, and its up-sampling module.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Disentangling Architecture and Training for Optical Flow

Sun¹,

Herrmann²,

Reda³

et al. 2022

Preprint

View full text Add to dashboard Cite

How important are training details and datasets to recent optical flow models like RAFT? And do they generalize? To explore these questions, rather than develop a new model, we revisit three prominent models, PWC-Net, IRR-PWC and RAFT, with a common set of modern training techniques and datasets, and observe significant performance gains, demonstrating the importance and generality of these training details. Our newly trained PWC-Net and IRR-PWC models show surprisingly large improvements, up to 30% versus original published results on Sintel and KITTI 2015 benchmarks. They outperform the more recent Flow1D on KITTI 2015 while being 3× faster during inference. Our newly trained RAFT achieves an Fl-all score of 4.31% on KITTI 2015, more accurate than all published optical flow methods at the time of writing. Our results demonstrate the benefits of separating the contributions of models, training techniques and datasets when analyzing performance gains of optical flow methods. Our source code will be publicly available.

show abstract

Section: Benchmark Resultsmentioning

confidence: 78%