Datasets:
File size: 3,165 Bytes
45741a4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | ---
license: cc-by-sa-4.0
pretty_name: ZDPShift
tags:
- stereo-matching
- disparity
- negative-disparity
- ZDP
- blender
- open-movies
task_categories:
- depth-estimation
size_categories:
- 10K<n<100K
---
# ZDPShift: Beyond the Zero-Disparity Plane in Stereo
<p align="center">
<img src="images/teaser.png" alt="Same scene rendered at five ZDP shifts; anaglyph + GT disparity + error map" width="100%"/>
</p>
Every public stereo benchmark assumes positive disparity values`d = fB/Z ≥ 0`. Mordern stereoscopic display — cinema 3D, VR, HMDs — actively uses `d < 0`. ZDPShift bridges the gap: the same artist-authored open-movie content rendered at **five ZDP shifts** Δ ∈ {−16, 0, +16, +24, +32} pixels, giving you a controlled continuum from textbook-positive to substantially-crossed disparities, with analytical ground truth at every pixel.
## News:
- [1st July] A technical report is on the way!
## The problem in one plot
<p align="center">
<img src="images/sota_failure.png" alt="EPE of public SOTA stereo models vs ZDP shift" width="80%"/>
</p>
Every public SOTA fails by **19–60×** the instant disparity crosses zero. ZDPShift gives you the data to fix that.
## What you get
| Split | Source | Scenes | Frames | Pairs | Size |
|---|---|---:|---:|---:|---:|
| `train/` | 6 Blender open movies | 30 | 4,134 | 20,670 | 318 GB |
| `test/` | Caminandes Llamigos | 4 | 271 | 1,355 | 35 GB |
Per pair: **left.png** + **right.png** (1920×1280 open-movie) + **disparity.npy** (signed float32) + **meta.json** (Δ, focal, baseline, K_L, K_R, % crossed).
## Gallery
<p align="center">
<img src="images/gallery.png" alt="ZDPShift scene gallery across the 8 source franchises" width="100%"/>
</p>
## Geometry
```
d(Z, Δ) = f · B / Z − Δ baseline B = 0.10 m, Δ ∈ {−16, 0, +16, +24, +32} px
```
Cameras stay physically parallel — only the principal points shift by ∓Δ/2. Multi-Δ self-consistency `d(Δᵢ) − d(Δⱼ) = Δⱼ − Δᵢ` holds to a median 0.04 px residual (sub-pixel Cycles noise).
```
rendered_datasets/<split>/<scene>/frame_<N>_<cam>/shift_{-16,+0,+16,+24,+32}/
├── left.png
├── right.png
├── disparity.npy
└── meta.json
```
## Render protocol
To keep ground truth well-defined at every pixel, three artist settings are overridden:
- **Aperture forced to f/64** — defocus blur kept below sub-pixel
- **Motion blur off** — fast objects have crisp per-frame correspondences
- **Compositor bypassed** — no grading / bloom / flare distortion after correspondence
Everything else (lighting, materials, scene composition, camera path, focal length per shot) is inherited from the original `.blend` unchanged. Cycles 32 spp + OPTIX denoising; sampler noise floor ≈ 0.005/255 per pixel.
## License
**CC-BY-SA-4.0**, inherited from the Blender Studio open movies. Redistribute and adapt with attribution and share-alike. |