Journal of Applied Optics, Volume. 46, Issue 2, 355(2025)
Fast optical flow estimation algorithm for edge GPU devices
Figures & Tables(9)
Fig. 1. Structure diagram of MiniFlow optical flow estimation neural network model
Fig. 3. Visualization effects of feature maps after channel rearrangement layer
Fig. 6. Visualization results of models on DAVIS
Table 1. Performance comparison on Sintel and KITTI datasets
View tableView in ArticleTable 1. Performance comparison on Sintel and KITTI datasets
方法 Sintel(train) KITTI-15(train) 参数/Mb 时间/s Clean Final Fl-epe F1-all/% RAFT 1.43 2.71 5.04 17.4 4.80 0.327* RAFT-ft (0.76) (1.22) (0.63) (1.5) 4.80 0.327* FlowNet2 2.02 3.14 10.06 30.37 162.52 0.116 FlowNet2-ft (1.45) (2.01) (2.30) (8.61) 162.52 0.116 LiteFlowNet 2.48 4.04 10.39 28.50 5.37 0.055 LiteFlowNet-ft (1.35) (1.78) (1.62) (5.58) 5.37 0.055 SPyNet 4.12 5.57 − − 1.20 0.050 SPyNet-ft (3.17) (4.32) − − 1.20 0.050 RAFT-small-10 2.21 3.38 9.55 28.67 0.99 0.041* PWC-Net 2.55 3.93 10.35 33.67 8.75 0.034 PWC-Net-ft (2.02) (2.08) (2.16) (9.80) 8.75 0.034 LiteFlowNetX 3.58 4.79 15.81 34.90 0.90 0.030 RAFT-small-4 2.89 4.22 11.77 34.23 0.99 0.027* PWC-Net-small 2.83 4.08 − − 4.08 0.024 FastFlowNet 2.89 4.14 12.24 33.10 1.37 0.011 FastFlowNet_v2 2.89 4.14 12.24 33.10 1.37 0.012* FastFlowNet-ft (2.08) (2.71) (2.13) (8.21) 1.37 0.011 MiniFlow 2.52 3.77 12.98 39.73 0.54 0.009* MiniFlow-ft （1.46） (1.88) （1.47） (6.09) 0.54 0.009*
Table 2. Performance comparison of RAFT-small, FastFlowNet_v2, and MiniFlow on NX
View tableView in ArticleTable 2. Performance comparison of RAFT-small, FastFlowNet_v2, and MiniFlow on NX
方法 Sintel (train) 速度/(frame/s) 参数/Mb Clean Final NX RAFT-small-4 2.89 4.22 2.1 0.99 FastFlowNet_v2 2.89 4.14 2.77 1.37 MiniFlow 2.52 3.77 12.57 0.54
Table 3. Performance comparison of different models on Sintel dataset
View tableView in ArticleTable 3. Performance comparison of different models on Sintel dataset
方法 Sintel (train) 时间/ms 参数/Mb Clean Final 1080Ti LMAC-Net 3.10 4.18 8.7 0.54 ResNet 3.32 4.46 10 0.56 FIUM 3.10 4.18 8.7 0.54 ConvGRU 3.08 4.39 8.9 0.82
Tools
Get Citation
Copy Citation Text
Ke SHI, Suzhen NIE, Dongxing LI, Jie CAO, Yunlong SHENG, Bin YAO, Honglin CHEN. Fast optical flow estimation algorithm for edge GPU devices[J]. Journal of Applied Optics, 2025, 46(2): 355
Paper Information
Category:
Received: Nov. 23, 2023
Accepted: --
Published Online: May. 13, 2025
The Author Email: Dongxing LI (李东兴)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20