Research map / 3D vision / Depth and stereo
Quanta Neural Networks: From Photons to Perception
Varun Sundar, Tianyi Zhang, Sacha Jungerman, Mohit Gupta
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add intermediate learnable QNN layers beyond the sensor-proximal layer (ablation) | QNN (adaptive integrator) · Depth estimation, dense compute | threshold accuracy (%): 53.2 -> 71.2 | Improved |
| add intermediate learnable QNN layers beyond the sensor-proximal layer (ablation) | QNN (adaptive integrator) · Point tracking, dense compute | average Jaccard (%): 60.1 -> 75.9 | Improved |
| add intermediate learnable QNN layers beyond the sensor-proximal layer (ablation) | QNN (adaptive integrator) · Intensity restoration i2k, dense compute | PSNR (dB): 30.3 -> 36.7 | Improved |
| use adaptive (Bayesian run-length) integration instead of fixed exponential smoothing (ablation) | QNN with intermediate layers · Intensity restoration, eventful top-10% | PSNR (dB): 32.7 -> 34.7 | Improved |
| use adaptive instead of fixed integration with intermediate layers, point tracking dense (ablation) | QNN with intermediate layers · Point tracking, dense compute | average Jaccard (%): 76.1 -> 75.9 | No clear effect |
| use adaptive instead of fixed integration without intermediate layers (ablation) | QNN sensor-proximal only · Depth estimation, dense compute | threshold accuracy (%): 50.8 -> 53.2 | Improved |
| run network incrementally with eventful computation (top-10%) instead of dense | QNN with adaptive integrator · Depth estimation | threshold accuracy (%): 71.2 -> 65.5 | Mixed |
| replace reconstruction-then-depth pipeline with QNN dense | DepthAnything-v2 QNN · Stress-ball depth | latency: 6000 ms -> 38 ms | Improved |
| add eventful computation on top of dense QNN | DepthAnything-v2 QNN · Stress-ball depth | GFLOP per output: 78 -> 26 | Improved |
| use change-driven sampling instead of uniform inference at comparable cost | QNN · Balloon burst | output fidelity / runtime: 150x run-time speedup | Improved |
| min-pool run-length estimates with 5x5 kernel | adaptive integrator · qualitative (Figure 3) | motion responsiveness | Improved |
Closest papers by meaning
- Streaming Quanta Sensors for Online, High-Performance Imaging and Vision (TPAMI 2024 · similarity 0.91)
- bit2bit: 1-bit quanta video reconstruction via self-supervised photon prediction (NeurIPS 2024 · similarity 0.87)
- Generative Quanta Color Imaging (CVPR 2024 · similarity 0.87)
- Towards 3D Vision with Low-Cost Single-Photon Cameras (CVPR 2024 · similarity 0.85)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.85)
- Quantization without Tears (CVPR 2025 · similarity 0.84)
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.
