TinyFlow improves edge optical flow without costly feature warping
TinyFlow: Breaking the Warping Bottleneck for Edge Optical Flow
RoboticsComputer Vision and Pattern Recognition
Summary
Optical flow is a way for computers to track movement between video frames, but the usual methods are too heavy for small devices because they rely on complicated steps called warping. The authors created TinyFlow, a lightweight network that skips this costly warping step by using a fixed-window correlation, making it easier to run on resource-limited devices. They carefully adjusted the model to work well with low-precision math, achieving near the same accuracy as bigger models while using fewer resources. Their model also showed better performance in tracking camera movement through video sequences on several test sets.
What this means in practice
- •For embedded system developers: Run efficient optical flow on small edge devices without complex data-dependent memory access steps to improve real-time motion tracking.
- •For drone navigation engineers: Use low-precision lightweight optical flow models to improve monocular visual odometry accuracy with limited onboard computing resources.
Authors
Xiao'ao Song, Chahat Deep Singh
Abstract
High-performance optical flow networks are difficult to deploy on resource-constrained edge accelerators because their iterative refinement warps features by the current flow, which relies on flow-dependent resampling and data-dependent memory access. We present TinyFlow, a 1.26M-parameter warp-free optical flow network that preserves iterative coarse-to-fine refinement by replacing the warp with a fixed-window correlation over constant offsets, so the whole network compiles as a single static INT8 graph. Quantization-aware fine-tuning, warp-teacher distillation, and model souping reduce the Float32-to-INT8 penalty to below 0.2 end-point error (EPE) on MPI-Sintel. TinyFlow outperforms EdgeFlowNet and NanoFlowNet on Sintel Clean, Sintel Final, and FlyingChairs in both Float32 and INT8, while using 54% fewer parameters than EdgeFlowNet, and our INT8 model achieves 6.29/7.13 EPE on Sintel Clean/Final. In flow-only monocular visual odometry on TartanAir and UZH-FPV, the same INT8 model yields the lowest trajectory error on seven of eleven sequences.