LibreUniMatchl-flow-sintel

Sintel finetune of LibreUniMatchl-flow. GMFlow+ scale2 with 6 local regression refinement steps, finetuned on MPI Sintel, repackaged for LibreYOLO's flow task.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreUniMatchl-flow-sintel.pt")
r = model(["frame1.jpg", "frame2.jpg"])[0]
r.flow.data      # (H, W, 2) float32, (dx, dy) in pixels on the original canvas
r.flow.to_rgb()  # Middlebury colour wheel

Source

Derived from autonomousvision/unimatch at commit 943a99006950ab6c84ea72a72beb147fdd03c19a, checkpoint gmflow-scale2-regrefine6-sintelft-6e39e2b9.pth from the upstream MODEL_ZOO.md. Copyright (c) 2022 autonomousvision. Code licensed under the MIT License.

Paper: Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao and Andreas Geiger. "Unifying Flow, Stereo and Depth Estimation." IEEE TPAMI 2023.

Modifications

Metadata wrap only: the upstream tensors are stored unchanged inside the LibreYOLO checkpoint schema (task flow). Learned parameters are unchanged. See weights/convert_unimatch_weights.py in the LibreYOLO source repository.

License

Upstream publishes these checkpoints without a separate weight licence. They are hosted here under the MIT licence of the releasing project; that grant is implied by the releasing project and is LibreYOLO's disclosed interpretation, not one confirmed by the publisher. See the LICENSE and NOTICE files in this repository.