LibreNeuFlowv2s-flow-things

NeuFlow v2 optical flow weights (the FlyingThings3D training stage (before Sintel / KITTI fine-tuning)), repackaged for LibreYOLO's flow task. NeuFlow v2 is a 9M-parameter flow network built for edge devices.

Training-data variant things; the canonical checkpoint is LibreNeuFlowv2s-flow.

Source

Weights derived from neuflow_things.pth in neufieldrobotics/NeuFlow_v2 at commit 204b5e3744461d90303b9ff82caa7a1bb56a2ca2. Copyright (c) 2024 Zhiyong Zhang, Aniket Gupta, Huaizu Jiang, Hanumant Singh. Paper: "NeuFlow v2: Push High-Efficiency Optical Flow To the Limit", arXiv:2408.10161.

Modifications

State-dict metadata-wrap only: keys and learned parameters are unchanged; the checkpoint is wrapped in the LibreYOLO v1.0 schema (task=flow, nc=1, names={0: "flow"}, variant=things, canvas 432x768). On an RTX 3090 in fp32 the LibreYOLO model reproduces the upstream NeuFlow module bit for bit (max absolute difference 0.0 px). See weights/convert_neuflowv2_weights.py and weights/parity_neuflowv2.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreNeuFlowv2s-flow-things.pt")
results = model(["frame_000.png", "frame_001.png"])  # N frames in, N - 1 flows out
flow = results[0].flow.data        # (H, W, 2) float32, (dx, dy) pixels, original canvas
rgb = results[0].flow.to_rgb()     # Middlebury colour wheel

License

Apache License 2.0. Upstream commits this checkpoint in its Apache-2.0 repository without a separate weight licence; LibreYOLO redistributes it under that repository licence, an implied grant from the releasing project and LibreYOLO's disclosed interpretation, not one confirmed by the authors. See the LICENSE and NOTICE files in this repository.