Skip to content
WorldAtlas WorldAtlas WorldAtlas
Home

Motion Accuracy

Motion Accuracy measures foreground motion salience inside supervised dynamic regions. SAM2 propagates initial segmentation masks through the prediction; SEA-RAFT estimates optical flow between consecutive frames. The metric tests whether masked object regions exhibit stronger motion than the background—it does not compare realized motion to a target trajectory, direction, or velocity profile.

For consecutive frame pair (It,It+1)(I_t, I_{t+1}), let ut\mathbf{u}_t be the optical-flow field and MtM_t the propagated (dilated) object mask. Define magnitude mt=∥ut∥2m_t = \|\mathbf{u}_t\|_2 and

Δt=max⁡p∈Mtmt(p)−max⁡p∉Mtmt(p)(1)\Delta_t = \max_{p \in M_t} m_t(p) - \max_{p \notin M_t} m_t(p) \tag{1}

The raw score is the mean salience over pairs whose propagated mask has both a nonempty foreground and a nonempty background:

Smotraw=1∣T∣∑t∈TΔt(2)S_{\mathrm{mot}}^{\mathrm{raw}} = \frac{1}{|T|}\sum_{t \in T} \Delta_t \tag{2}

WorldAtlas applies z-score calibration with μ=−0.1965\mu = -0.1965, σ=2.3687\sigma = 2.3687. The anchors [−1.6147,1.5180][-1.6147, 1.5180] map linearly to [0.25,0.75][0.25, 0.75]; values outside the anchors are extrapolated, then clipped to [0,1][0,1]. Raw zero maps to approximately 0.52096. Higher scores indicate stronger foreground flow relative to background, not action/direction accuracy or mask annotation accuracy.

Protocol aligned_mask_contrast_v2 requires verified image/video correspondence, mask metadata matching the source, and an explicit conditioning resize/crop transform. Source masks are aligned to the decoded first frame before SAM2 propagation. Adjacent decoded frame pairs with both regions nonempty use the official max-minus-max contrast. Missing masks or non-finite flow remain errors. An all-foreground or empty/all-background propagated mask skips that pair; if every pair skips, the sample is N/A. Skips are never a fabricated zero and are excluded from the suite mean. Explicit manual deletion of all source instances is also N/A. Reports include pair coverage, skipped-pair counts, and foreground fractions. This protocol is not interchangeable with historical scores.

The max-minus-max statistic is sensitive to flow outliers and has units of decoded pixels per frame. Native resolution/FPS differences can affect model comparisons. These are limitations of the retained metric, not properties fixed by refined masks; a robust percentile or resolution/FPS-normalized variant would need separate calibration and a new metric name.

Verified Same-Input Examples

Section titled “Verified Same-Input Examples”

These current-protocol examples contrast foreground motion against a fixed-camera target. A higher score does not establish that the object followed the requested direction or action. Audit manifest: docs/assets/metric_examples/motion_accuracy_current_cases_20260927.json.

Input image: image_dynamic_stylized_articulated_0001
Input image
Input verified · same benchmark case · gap 0.8942

image_dynamic_stylized_articulated_0001

Suite: image_dynamicScene: stylized / articulated
Benchmark prompt target

The red-and-white cow grazes in the bright pastoral field, lowering its head to the grass and making small chewing movements. Flowers, rolling hills.

prompt source: prompt_target

The cow lowers its head while the field stays comparatively still; the lower-scoring clip shifts more of the scene. Both first frames passed geometric matching against the displayed input; scores compare this metric only.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 0.8942

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.0000
Input image: image_dynamic_photorealistic_rigid_0001
Input image
Input verified · same benchmark case · gap 0.8178

image_dynamic_photorealistic_rigid_0001

Suite: image_dynamicScene: photorealistic / rigid
Benchmark prompt target

A long freight train rolls along the railroad tracks through dry grass, its boxcars translating together across the open landscape.

prompt source: prompt_target

The train moves against a steadier hillside; the lower-scoring clip also moves the background. Both first frames passed geometric matching against the displayed input; scores compare this metric only.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 0.8178

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.0000
Input image: image_dynamic_stylized_fluid_0001
Input image
Input verified · same benchmark case · gap 1.0000

image_dynamic_stylized_fluid_0001

Suite: image_dynamicScene: stylized / fluid
Benchmark prompt target

A forest river flows gently between rocky banks, its surface moving downstream and catching warm sunlight.

prompt source: prompt_target

The river surface moves while the banks stay comparatively fixed; the lower-scoring clip moves through the whole scene. Both first frames passed geometric matching against the displayed input; scores compare this metric only.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 1.0000

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.0000

References

Section titled “References”