# DPFlow
DPFlow is an adaptive optical flow estimator by Henrique Morimitsu and co-authors (CVPR 2025). Its dual-pyramid encoder adapts its depth to the input size, so a model trained at low resolution generalises to 4K and 8K frames. LibreYOLO ports it from ptlflow; the weights are licensed CC BY-NC-SA 4.0, for non-commercial use only.
Tasks: Optical flow. Install: pip install libreyolo.
Verified against LibreYOLO v1.8.0.

DPFlow chooses the number of pyramid levels from the image diagonal:
three levels around 1K, one more per doubling. LibreYOLO keeps that rule and
runs at the native frame resolution by default, which is where the adaptive
depth matters. Pick it for research on high-resolution footage; its weights
are non-commercial.

## Install

DPFlow 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 `LibreDPFlow` is also exported from `libreyolo` if you prefer to build the
model directly.

**Python**

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreDPFlowm-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("LibreDPFlowm-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=LibreDPFlowm-flow.pt source=frame_0001.jpg,frame_0002.jpg save=true
libreyolo predict model=LibreDPFlowm-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. DPFlow 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, size `m`, with four checkpoints:

| File | Training stage |
| --- | --- |
| `LibreDPFlowm-flow.pt` | FlyingThings3D (`things`), the generalisation checkpoint (canonical) |
| `LibreDPFlowm-flow-sintel.pt` | Sintel fine-tune |
| `LibreDPFlowm-flow-kitti.pt` | KITTI fine-tune |
| `LibreDPFlowm-flow-spring.pt` | Spring fine-tune |

The iteration schedule is fixed at upstream's test setting, 4 refinement
iterations per pyramid level. Every pair is estimated independently: the
upstream warm start across frames is not used. `imgsz` caps the long side
(default 8192, in practice no cap); frames above it are downscaled and the
field is scaled back to the original canvas.

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("LibreDPFlowm-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=LibreDPFlowm-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 |
| --- | --- | --- | --- |
| `LibreDPFlowm-flow.pt` |  | Optical flow | cc-by-nc-sa-4.0 |
| `LibreDPFlowm-flow-sintel.pt` |  | Optical flow | cc-by-nc-sa-4.0 |
| `LibreDPFlowm-flow-kitti.pt` |  | Optical flow | cc-by-nc-sa-4.0 |
| `LibreDPFlowm-flow-spring.pt` |  | Optical flow | cc-by-nc-sa-4.0 |

## 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: DPFlow, Henrique Morimitsu (ptlflow)
- Upstream license: CC-BY-NC-SA-4.0
- Upstream source: https://github.com/hmorimitsu/ptlflow
- LibreYOLO code: Apache-2.0
- Weights: CC-BY-NC-SA-4.0, republished at https://huggingface.co/LibreYOLO
- Interpretation: The code and the weights carry different licenses. The weights were trained within ptlflow and are released under CC BY-NC-SA 4.0: commercial use is not allowed, redistribution needs attribution, and anything you build from the weights and share must carry the same license; the DPFlow README adds that the weights are for academic and research purposes only, because of the training datasets' licenses. Treat all four checkpoints (things as the canonical m, and the sintel, kitti and spring fine-tunes) as research weights. The code is Apache-2.0: ptlflow's models/dpflow at commit 43cdc2b5 is vendored under libreyolo/models/dpflow/_vendor/, with timm-derived helpers (Apache-2.0) and an input padder adapted from RAFT (BSD-3-Clause); LibreYOLO's wrapper code is MIT. The weights come from the ptlflow weights1 release and are mirrored on the LibreYOLO Hugging Face org with the license text and a non-commercial banner, without gating. Training is out of scope.

DPFlow is the work of Henrique Morimitsu and co-authors, released in
[ptlflow](https://github.com/hmorimitsu/ptlflow). The code and the weights
carry different licences.

- The code is **Apache-2.0**: ptlflow's `models/dpflow` is vendored with
  timm-derived helpers (Apache-2.0) and an input padder adapted from RAFT
  (BSD-3-Clause).
- The weights are **CC BY-NC-SA 4.0**: **non-commercial use only**,
  attribution required, and anything you build from them and share must carry
  the same licence. The DPFlow README adds that the weights are for academic
  and research purposes only, because of the training datasets' licences.
  Treat all four checkpoints as research weights.

LibreYOLO mirrors the weights with the licence text and a non-commercial
banner, without gating, and prints a notice when they load.

## Citation

```bibtex
@InProceedings{Morimitsu2025DPFlow,
  author    = {Morimitsu, Henrique and Zhu, Xiaobin and Cesar-Jr., Roberto M. and Ji, Xiangyang and Yin, Xu-Cheng},
  booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  title     = {{DPFlow}: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework},
  year      = {2025},
}
```

Copied from https://github.com/hmorimitsu/ptlflow/tree/43cdc2b5fc7228e7090009a66e8e28e0b48e31ba/ptlflow/models/dpflow#citation
