Research map / 3D vision / Depth and stereo
All-directional Disparity Estimation for Real-world QPD Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DPNet and QuadNet for dual-pixel and quad-pixel (QPD) disparity estimation with an illumination-invariant module, a sub-pixel coarse-to-fine module, and an edge-aware fusion of horizontal and vertical disparities, plus a new 2,100-image real QPD disparity dataset QPD2K.
Problem
Disparity estimation from real-world QPD sensors is hard due to uneven sub-view illumination, a narrow baseline yielding tiny disparities, and the need to fuse two-directional disparity; existing QPD/DP datasets lack true disparity ground truth.
Approach
Uses differential (edge-based) convolutions for illumination-invariant features, a sub-pixel cost volume built by shifting warped features around a coarse disparity (RAFT-stereo style iterations), and an edge-aware fusion that weights horizontal and vertical disparities by directional edge maps, followed by census-based refinement.
Key results
- QuadNet reaches bad 0.5 of 0.114 and EPE of 0.201 on QPD2K, versus 0.189 and 0.272 for the best prior baseline DLNR.
- DPNet achieves bad 0.5 of 0.160 and EPE of 0.238 on QPD2K, beating DLNR (0.189 / 0.272).
- On DP5K, DPNet's AI(1) error is 0.034 versus 0.078 for the next-best SFBD, and DPNet outperforms the next-best method on DP-disp by about 37.5%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DPNet and fuse horizontal and vertical disparity with QuadNet | QPD disparity estimation · QPD2K | epe: 0.272 -> 0.201 | Helped |
| Use DPNet for disparity estimation | DP disparity estimation · DP5K | AI(1): 0.078 -> 0.034 | Helped |
| Use DPNet for disparity estimation | DP disparity estimation · DP-disp | AI(1): 0.041 -> 0.025 | Helped |
| Add illumination-invariant, sub-pixel, and fusion modules to Base (ablation) | RAFT-stereo baseline on QPD2K · QPD2K | epe: 0.353 -> 0.201 | Helped |
| Add illumination-invariant and sub-pixel modules to Base (ablation) | RAFT-stereo baseline on QPD2K · QPD2K | epe: 0.353 -> 0.238 | Helped |
| Add illumination-invariant module to Base (ablation) | RAFT-stereo baseline on QPD2K · QPD2K | epe: 0.353 -> 0.264 | Helped |
| Add sub-pixel module to Base (ablation) | RAFT-stereo baseline on QPD2K · QPD2K | epe: 0.353 -> 0.259 | Helped |
| Use 4 differential convolutions instead of CDC+ADC (ablation) | DPNet illumination-invariant module · QPD2K | epe: 0.287 -> 0.264 | Helped |
| Replace differential convolutions with vanilla convolution (ablation) | DPNet illumination-invariant module · QPD2K | epe: 0.264 -> 0.331 | Hurt |
| Use bilinear sampling for sub-pixel shifts rather than nearest sampling (ablation) | DPNet cost volume · QPD2K | epe: 0.289 -> 0.259 | Helped |
| Use bilinear sampling for sub-pixel shifts rather than phase sampling (ablation) | DPNet cost volume · QPD2K | epe: 0.273 -> 0.259 | Helped |
| Use edge-aware fusion instead of convolution fusion (ablation) | QuadNet · QPD2K | epe: 0.211 -> 0.201 | Helped |
Surprising findings
- Indirect disparity estimation via a center reference image performs poorly on QPD because the limited disparity range makes estimating only half the disparity insufficient.
- Self-supervised PSF-symmetry methods (SFBD, DPdisp) underperform supervised stereo-style methods on real QPD data, attributed to PSF asymmetry under noise and uneven illumination.
Limitations
- Ground truth disparity is derived from structured-light depth and calibration, which is costly and limited to scenes suited to structured light.
- The disparity range is small (about -4 to +4 pixels), so methods designed for large-disparity stereo do not transfer directly.
Experimental setup
Datasets: QPD2K DP5K DP-disp
Benchmarks: QPD2K DP5K DP-disp
Metrics: bad 0.3 bad 0.5 bad 1 end-point error affine-invariant error Spearman correlation geometric mean
Baselines: DPdisp SFBD CCA IEGV-Stereo QPDNet Mc-stereo S-RAFT S-IEGV DLNR FaceDPNet DDDNet WDDNet
Builds on: RAFT-stereo DPdisp Kim's SFBD kernel-split symmetry loss Census transform
Closest papers by meaning
- Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model (CVPR 2025 · similarity 0.84)
- QuartDepth: Post-Training Quantization for Real-Time Depth Estimation on the Edge (CVPR 2025 · similarity 0.84)
- Adaptive Fusion of Single-View and Multi-View Depth for Autonomous Driving (CVPR 2024 · similarity 0.83)
- Dual Exposure Stereo for Extended Dynamic Range 3D Imaging (CVPR 2025 · similarity 0.83)
- Differentiable Display Photometric Stereo (CVPR 2024 · similarity 0.83)
- HELVIPAD: A Real-World Dataset for Omnidirectional Stereo Depth Estimation (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.