Six state-of-the-art stereo matching models collapse the moment disparity crosses zero, their error rising by 4.6–37×.
Stereoscopic content, from cinema 3D to VR, lives there.

Video

Unseen dataset (with GT): Enables negative disparities. Preserves performances on positive-only datasets.

In the wild: Making stereo matching on real-world 3D movies possible.

Overview

One frame, five ZDP shifts

One frame rendered at five positions of the zero-disparity plane; only the plane moves. Released FoundationStereo never leaves the positive side and returns a plausible disparity map of the wrong sign; after fine-tuning on horizontally translated SceneFlow, never on this benchmark, the same backbone follows the ground truth across the whole signed range.

FoundationStereo across ZDP shifts, zero-shot vs ours.

The blind spot

Every backbone breaks at zero

Six released backbones, three image and three video, degrade the instant disparity crosses zero: EPE reaches 13–150 px at Δ=+32 on identical scene content (dashed). Fine-tuned on translated SceneFlow only (solid), all six stay within 2.9–5.2 px across the entire signed range.

The dataset

ZDPShift

4,275 open-movie frames rendered at five zero-disparity-plane shifts each, with analytical signed ground-truth disparity. Evaluation-only: no released model trains on it.

21,375

stereo pairs

4,275

frames × 5 ZDP shifts

scenes / 7 open movies

signed

analytical GT disparity

Results

Flat across the full signed range

End-point error (px, native) at the hardest shift, Δ=+32, on the full benchmark of 21,375 pairs — 960-px inference. No model sees any benchmark data: training is horizontally translated SceneFlow only. Frozen keeps the pretrained matching features fixed and trains only the layers that read disparity out of them.

Model	Zero-shot	+ ours (full)	+ ours (frozen features)
RAFT-Stereo	149.7	4.68	4.89
IGEV-Stereo	44.9	3.97	4.06
FoundationStereo	75.3	2.90	3.01
DynamicStereo	20.8	5.19	5.19
BiDAStereo	36.2	4.49	4.51
StereoAnyVideo	13.0	3.98	4.11