Data Scaling Laws in Stereo Matching: From In-domain Fitting to Cross-domain Generalization

What it claims

XPENG Motors researchers (October 2026) train one fixed 11M-parameter stereo model on nested subsets of four synthetic datasets and their mixture. In-domain disparity error follows a power law and then saturates. The fitted exponents across the five curves fall between 0.216 and 0.363, over a range of roughly 2^30 to 2^38 labelled pixels. The transition point stays roughly fixed for a given model, and the authors introduce an "Effective Fitting Utilization" metric for data redundancy based on it. In zero-shot tests, such as KITTI12, more source data gives diminishing gains. The authors report that distribution and coverage diversity matter more than source-data scale.

How it works

XStereo dataset. XStereo is rendered in Unreal Engine 5 from 162 indoor and outdoor scenes, with 200K stereo pairs at 512×960. It has the following properties:

• Focal length is 500 to 1100 px and the stereo baseline is 6 to 25 cm.
• Lighting, lens flare and fog are randomized.
• Cameras walk at human-like viewpoints.
• Translucent objects have depth labels.
• Each scene contributes at most 2,000 samples.
• Over-exposed and texture-less frames were removed by human QA.

Scenes and per-scene samples can be subsampled independently, which is what allows diversity and density to be studied separately.

Model. The model follows FoundationStereo with these changes:

• An adapted EfficientNet replaces the Depth Anything backbone and gives 128-channel features at 1/4 resolution.
• The cost volume is built from the element-wise product of left and right features, passed through a learnable 128→32 projector.
• Each 3×3×3 3D convolution is split into a 1×3×3 kernel followed by a 7×1×1 kernel.
• Training uses 8 GRU refinement steps, a smooth-L1 loss on the initial disparity and an exponentially weighted L1 loss (γ = 0.85) on the GRU outputs.

Scaling protocol.

• Subsets are nested (D1 ⊂ D2 ⊂ …). They start at 2,500 pairs (1,250 for Scene Flow) and double until the full set.
• Data scale is measured as unique pixels with ground truth, not as image count. Compute is measured as accumulated processed GT pixels.
• Compute is fixed rather than epochs: 45K steps per run, and 100K steps for mixed training.
• The optimizer is AdamW with a cosine schedule, 10% warm-up and a peak learning rate of 5e-4. These settings were tuned on a small mixed subset.
• Maximum disparity is 128. Samples with less than 30% GT in the range [0, 128] are dropped, and no data augmentation is used.

Evidence

Training data.

• XStereo: 200K pairs.
• Scene Flow: 40K pairs.
• FSD: 290K pairs (partial).
• TartanAir: 110K pairs (partial).
• FallingThings: 30K pairs, used only in the mix.
• Held-out test sets are 5K pairs for Scene Flow and FallingThings and 20K pairs for the others.

Zero-shot targets. KITTI12 (194 samples), KITTI15 (200), RSRD (2,789), DrivingStereo (20K) and WHUStereo (1,618). Middlebury and ETH3D were excluded as too small.

Metrics. EPE, RMSE, BP-X and KITTI D1.

The fetched text ends partway through the implementation details. Per-dataset tables, the cross-domain numbers and the mapping of each exponent to its dataset are not in the extracted text, beyond the Figure 1 values.

How to use it

XStereo is published at https://huggingface.co/datasets/Tong-Zhao/XStereo. The page states no licence, and no code or model weights are mentioned.

Limits

• The authors call the study preliminary.
• Model size and architecture are fixed throughout, so the saturation point is specific to this one 11M model. Model scaling is not tested.
• The cap of disparity 128 and the lack of augmentation differ from typical production training.
• FSD and TartanAir were only partially used.
• The Effective Fitting Utilization metric is named but not defined in the fetched text.