# MEMFOF
MEMFOF is a memory-efficient multi-frame optical flow network from the MSU Graphics & Media Lab (ICCV 2025 Highlight). It reads three frames, t-1, t and t+1, and predicts the flow of the middle frame, aimed at Full HD video. It is the one flow family in LibreYOLO that reads a three-frame window.
Tasks: Optical flow. Install: pip install libreyolo.
Verified against LibreYOLO v1.8.0.

MEMFOF is a RAFT descendant that predicts the backward (`t` to
`t-1`) and forward (`t` to `t+1`) flow of the middle frame in one pass.
LibreYOLO reports the forward flow as `Results.flow`; the backward flow is
computed but kept out of the task contract for now. The runner hands it the
`(t-1, t, t+1)` window and repeats frame 0 as `t-1` at the start of a
sequence, so you still pass an ordinary list, folder or video. Pick it for
real-world video at high resolution.

## Install

MEMFOF needs no optional extra. Everything it imports is in the base install.

```bash
pip install libreyolo
```

## Predict

Weights download from the LibreYOLO Hugging Face organization on first use.
The class `LibreMEMFOF` is also exported from `libreyolo` if you prefer to build the
model directly.

**Python**

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreMEMFOFm-flow.pt")   # weights download on first use
results = model(["frame_0001.jpg", "frame_0002.jpg"])

flow = results[0].flow
print(flow.data.shape)        # (H, W, 2) float32, (dx, dy) in pixels
print(flow.mean_magnitude)    # mean displacement in pixels
```

**Video**

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreMEMFOFm-flow.pt")

# N frames give N - 1 results; result i is the flow from frame i to i + 1.
for result in model("video.mp4", stream=True, save=True, save_flow=True):
    print(result.flow.max_magnitude)
```

**CLI**

```bash
libreyolo predict model=LibreMEMFOFm-flow.pt source=frame_0001.jpg,frame_0002.jpg save=true
libreyolo predict model=LibreMEMFOFm-flow.pt source=video.mp4 save=true arrows=true save_flow=true
```

Flow needs at least two frames: a single image raises. A list of images, a
directory sorted by name, a video or a stream is one sequence, and `N` frames
give `N - 1` results. `result.flow` is a `FlowField` with an `(H, W, 2)` float32
array on the original canvas of the first frame; `flow.data[y, x]` is
`(dx, dy)` in pixels, `+x` right and `+y` down. MEMFOF fills no confidence map, so `flow.confidence` is `None`. `result.boxes` is
`None`. `save=True` writes the Middlebury colour wheel, `arrows=True` or
`warp=True` pick the other views, and `save_flow=True` also writes each pair's
raw field as a `.flo` file. The [optical flow task](/docs/tasks/optical-flow)
page describes the payload and views in full.

## Variants

One architecture (ResNet-34 FPN), size `m`, with six checkpoints:

| File | Upstream checkpoint |
| --- | --- |
| `LibreMEMFOFm-flow.pt` | MEMFOF-Tartan-T-TSKH, recommended by the authors for real-world video (canonical) |
| `LibreMEMFOFm-flow-tartan-t.pt` | an earlier published training stage |
| `LibreMEMFOFm-flow-tartan.pt` | an earlier published training stage |
| `LibreMEMFOFm-flow-kitti.pt` | KITTI fine-tune |
| `LibreMEMFOFm-flow-sintel.pt` | Sintel fine-tune |
| `LibreMEMFOFm-flow-spring.pt` | Spring fine-tune |

Frames run at native resolution with upstream's replicate padding to a
multiple of 16. Frames whose long side exceeds `imgsz` (default 1920, MEMFOF's
Full HD target) are downscaled first and the field is scaled back. Inference is
fp32 with 8 refinement iterations, upstream's defaults.

Training is not offered: `train()` raises `NotImplementedError`.

## Validate

`val()` runs the shared flow validator on a flow dataset (frames per sequence
folder, Middlebury `.flo`, KITTI 16-bit PNG or `.npy` ground truth; see the
[dataset format](/docs/tasks/optical-flow#dataset-format)). It reports the mean
end-point error `metrics/epe`, `metrics/px1` and `metrics/px3` (share of valid
pixels under 1 and 3 px of error) and the KITTI outlier rate
`metrics/fl_all`.

**Python**

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreMEMFOFm-flow.pt")
metrics = model.val(data="my_flow.yaml")

print(metrics["metrics/epe"])      # mean end-point error, pixels
print(metrics["metrics/px1"])      # share of pixels under 1 px error
print(metrics["metrics/fl_all"])   # KITTI outlier rate
```

**CLI**

```bash
libreyolo val model=LibreMEMFOFm-flow.pt data=my_flow.yaml
```

## Export

| Task | ONNX | TorchScript | ExecuTorch | TensorRT | OpenVINO | Paddle | MNN | RKNN | Axelera | ncnn | TFLite | CoreML | Core AI |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Optical flow |  |  |  |  |  |  |  |  |  |  |  |  |  |

A "yes" means the export is supported. An empty cell means the exporter refuses that combination.

Export is not available yet for any format. Optical flow export is blocked
until the two-input graph (`image1`, `image2`) and the backend vector-rescale
contract of the flow task are implemented for a family.

## Checkpoints

Every published weight file for this family.

| File | Input (px) | Task | Weights license |
| --- | --- | --- | --- |
| `LibreMEMFOFm-flow.pt` |  | Optical flow | bsd-3-clause |
| `LibreMEMFOFm-flow-kitti.pt` |  | Optical flow | bsd-3-clause |
| `LibreMEMFOFm-flow-sintel.pt` |  | Optical flow | bsd-3-clause |
| `LibreMEMFOFm-flow-spring.pt` |  | Optical flow | bsd-3-clause |
| `LibreMEMFOFm-flow-tartan.pt` |  | Optical flow | bsd-3-clause |
| `LibreMEMFOFm-flow-tartan-t.pt` |  | Optical flow | bsd-3-clause |

## Licensing

Check the license on the Hugging Face repository of the specific weights you download. That repository is authoritative and licenses are not always uniform across a family. This is a description of the licenses involved, not legal advice.

- Original work: MEMFOF, MSU Graphics & Media Lab
- Upstream license: BSD-3-Clause
- Upstream source: https://github.com/msu-video-group/memfof
- LibreYOLO code: MIT
- Weights: BSD-3-Clause, republished at https://huggingface.co/LibreYOLO
- Interpretation: BSD-3-Clause is a permissive license: the code and weights can be used in commercial and closed-source products. Redistributions must keep the copyright notice, the license conditions and the disclaimer, and the authors' names may not be used to endorse derived products without permission. The upstream inference code at commit a51de9fc is vendored under libreyolo/models/memfof/_vendor/ with its BSD-3-Clause LICENSE; its gma.py attention modules come from GMA (zacjiang/GMA) under the WTFPL, shipped as LICENSE-GMA. LibreYOLO's wrapper code is MIT. The upstream checkpoints on Hugging Face (egorchistov/optical-flow-MEMFOF-*) are tagged bsd-3-clause but ship no license file, so LibreYOLO mirrors the six converted checkpoints (canonical, kitti, sintel, spring, tartan, tartan-t) under the code repository's BSD-3-Clause license. No gating; inference only.

MEMFOF is the work of the
[MSU Graphics & Media Lab](https://github.com/msu-video-group/memfof). The
upstream inference code is vendored under its BSD-3-Clause licence; its
attention modules come from GMA under the WTFPL. The upstream checkpoints on
Hugging Face are tagged bsd-3-clause but ship no licence file, so LibreYOLO
mirrors the six converted checkpoints under the code repository's BSD-3-Clause
licence, without gating. BSD-3-Clause allows commercial and closed-source use
as long as the copyright notice, conditions and disclaimer are kept.

## Citation

```bibtex
@article{bargatin2025memfof,
  title={MEMFOF: High-Resolution Training for Memory-Efficient Multi-Frame Optical Flow Estimation},
  author={Bargatin, Vladislav and Chistov, Egor and Yakovenko, Alexander and Vatolin, Dmitriy},
  journal={arXiv preprint arXiv:2506.23151},
  year={2025}
}
```

Copied from https://github.com/msu-video-group/memfof
