Event

TinyFlow: A Warp-Free Optical Flow Network Compilable into a Single INT8 Static Graph

1 reports · 1 sources

Xiao'ao Song and Chahat Deep Singh submitted a paper proposing TinyFlow, a 1.26M-parameter warp-free optical flow network. It replaces feature resampling that depends on the current optical flow with fixed-window, constant-offset correlation, while retaining coarse-to-fine iterative refinement, allowing the entire network to be compiled into a single static INT8 graph. The authors report that quantization-aware fine-tuning, warp teacher distillation, and model soup reduce the Float32-to-INT8 accuracy loss to within 0.2 EPE on MPI-Sintel; the INT8 model achieves 6.29/7.13 EPE on Sintel Clean/Final. In pure optical-flow monocular visual odometry on TartanAir and UZH-FPV, the same INT8 model had the lowest trajectory error on 7 of 11 sequences.