shijianjian commited on
Commit
45741a4
·
verified ·
1 Parent(s): 01217b6

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +79 -0
README.md ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-sa-4.0
3
+ pretty_name: ZDPShift
4
+ tags:
5
+ - stereo-matching
6
+ - disparity
7
+ - negative-disparity
8
+ - ZDP
9
+ - blender
10
+ - open-movies
11
+ task_categories:
12
+ - depth-estimation
13
+ size_categories:
14
+ - 10K<n<100K
15
+ ---
16
+
17
+ # ZDPShift: Beyond the Zero-Disparity Plane in Stereo
18
+
19
+ <p align="center">
20
+ <img src="images/teaser.png" alt="Same scene rendered at five ZDP shifts; anaglyph + GT disparity + error map" width="100%"/>
21
+ </p>
22
+
23
+ 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.
24
+
25
+ ## News:
26
+ - [1st July] A technical report is on the way!
27
+
28
+ ## The problem in one plot
29
+
30
+ <p align="center">
31
+ <img src="images/sota_failure.png" alt="EPE of public SOTA stereo models vs ZDP shift" width="80%"/>
32
+ </p>
33
+
34
+ Every public SOTA fails by **19–60×** the instant disparity crosses zero. ZDPShift gives you the data to fix that.
35
+
36
+ ## What you get
37
+
38
+ | Split | Source | Scenes | Frames | Pairs | Size |
39
+ |---|---|---:|---:|---:|---:|
40
+ | `train/` | 6 Blender open movies | 30 | 4,134 | 20,670 | 318 GB |
41
+ | `test/` | Caminandes Llamigos | 4 | 271 | 1,355 | 35 GB |
42
+
43
+ Per pair: **left.png** + **right.png** (1920×1280 open-movie) + **disparity.npy** (signed float32) + **meta.json** (Δ, focal, baseline, K_L, K_R, % crossed).
44
+
45
+ ## Gallery
46
+
47
+ <p align="center">
48
+ <img src="images/gallery.png" alt="ZDPShift scene gallery across the 8 source franchises" width="100%"/>
49
+ </p>
50
+
51
+ ## Geometry
52
+
53
+ ```
54
+ d(Z, Δ) = f · B / Z − Δ baseline B = 0.10 m, Δ ∈ {−16, 0, +16, +24, +32} px
55
+ ```
56
+
57
+ 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).
58
+
59
+ ```
60
+ rendered_datasets/<split>/<scene>/frame_<N>_<cam>/shift_{-16,+0,+16,+24,+32}/
61
+ ├── left.png
62
+ ├── right.png
63
+ ├── disparity.npy
64
+ └── meta.json
65
+ ```
66
+
67
+ ## Render protocol
68
+
69
+ To keep ground truth well-defined at every pixel, three artist settings are overridden:
70
+
71
+ - **Aperture forced to f/64** — defocus blur kept below sub-pixel
72
+ - **Motion blur off** — fast objects have crisp per-frame correspondences
73
+ - **Compositor bypassed** — no grading / bloom / flare distortion after correspondence
74
+
75
+ 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.
76
+
77
+ ## License
78
+
79
+ **CC-BY-SA-4.0**, inherited from the Blender Studio open movies. Redistribute and adapt with attribution and share-alike.