Research map / 3D vision / Depth and stereo
Focal Split: Untethered Snapshot Depth from Differential Defocus
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.
Builds Focal Split, a handheld, battery-powered snapshot depth camera that uses a beamsplitter and two defocused photosensors with a closed-form depth-from-differential-defocus algorithm running on a Raspberry Pi 5.
Problem
Existing depth-from-differential-defocus (DfDD) cameras capture defocused images sequentially, so they blur under scene motion, and they are tethered to wall power and laptops, limiting low-power untethered depth sensing.
Approach
A beamsplitter and two photosensors at different sensor distances capture two differently defocused images in one snapshot; the authors derive a pixel-wise depth formula using image derivatives after magnification-aligning the two images, costing about 500 FLOPs per pixel.
Key results
- Prototype outputs 480×360 sparse depth maps at 2.1 FPS on a Raspberry Pi 5 while consuming 4.9 W, powered by a 5 V 10,000 mAh battery.
- On dynamic scenes, Focal Split achieves 41.82 mm MAE versus 107.69 mm for Focal Track and 316.78 mm for Tang et al.
- The algorithm uses about 500 FLOPs per pixel and works over 0.4 m to 1.2 m, with working range rising from 860 mm to 940 mm at 40% sparsity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use snapshot capture of the image pair instead of sequential capture | Focal Split / DfDD · Front-parallel textured planes, dynamic scenes | MAE (mm): 107.69 -> 41.82 | Helped |
| Use snapshot capture of the image pair instead of sequential capture | Focal Split / DfDD · Front-parallel textured planes, dynamic scenes | working range (m): 0.295 -> 0.860 | Helped |
| Use snapshot capture rather than sequential capture | Focal Split / DfDD · Dynamic scenes | depth map quality | Helped |
| Discard the 40% least-confident pixels using the confidence threshold (ablation) | Focal Split · Front-parallel textured planes | working range: 860 mm -> 940 mm | Helped |
| Use the proposed aligned-image depth equation instead of the previously suggested unaligned-image equation (ablation) | Focal Split depth estimation · Synthetic data across sensor-distance variations | depth estimation error | Helped |
| Apply the proposed method directly to pillbox-shaped PSFs instead of Gaussian PSFs (ablation) | Focal Split depth estimation · Synthetic data with pillbox PSFs | depth prediction error | Hurt |
Limitations
- Depth estimates are sparse and inaccurate where both foreground and background have strong texture.
- Textureless regions yield no depth because the derivative ratio degenerates.
- Working range is limited to roughly 0.4 m to 1.2 m around the focal planes.
Experimental setup
Datasets: real captured images of front-parallel textured planes synthetic image pairs for sensitivity analysis
Benchmarks: real captured textured planes (static and dynamic scenes)
Metrics: mean absolute error (MAE) working range FPS power consumption (W)
Baselines: Focal Flow Tang et al. Focal Track COD
Builds on: depth from differential defocus Focal Track Focal Flow
Closest papers by meaning
- Passive Snapshot Coded Aperture Dual-Pixel RGB-D Imaging (CVPR 2024 · similarity 0.84)
- Multi-Focus Image Fusion via Explicit Defocus Blur Modelling (AAAI 2025 · similarity 0.83)
- Blurry-Edges: Photon-Limited Depth Estimation from Defocused Boundaries (CVPR 2025 · similarity 0.82)
- Learned Binocular-Encoding Optics for RGBD Imaging Using Joint Stereo and Focus Cues (CVPR 2025 · similarity 0.82)
- Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model (CVPR 2025 · similarity 0.81)
- SVDC: Consistent Direct Time-of-Flight Video Depth Completion with Frequency Selective Fusion (CVPR 2025 · similarity 0.81)
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.