Research map / 3D vision / Depth and stereo
ParamISP: Learned Forward and Inverse ISPs using Camera Parameters
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 ParamISP, a learned forward and inverse ISP pair that conditions on EXIF camera parameters via a ParamNet module, improving sRGB/RAW reconstruction and enabling downstream RAW deblurring and HDR applications.
Problem
Learned forward (RAW-to-sRGB) and inverse (sRGB-to-RAW) camera ISP models ignore the large variations in ISP behavior caused by camera parameters such as ISO, exposure time, aperture and focal length, and use simple architectures that limit RAW reconstruction quality.
Approach
Converts EXIF optical parameters into a feature vector via non-linear equalization (multiple nonlinear mappings normalized to [0,1]) with random dropout, and feeds it to LocalNet and GlobalNet; combines this with a non-learnable CanoNet for demosaicing, white balance and color transform, and a GlobalNet that iterates gamma correction and quadratic transforms.
Key results
- ParamISP improves sRGB-to-RAW PSNR on D7000 to 38.49 dB versus 35.52 dB for CycleISP with 0.7M params versus 3.1M.
- ParamISP achieves 34.14 dB RAW-to-sRGB PSNR on D7000 versus 30.20 dB for InvISP, and reports an average gain of 1.93 dB (RAW) and 1.84 dB (sRGB) over prior learned ISPs.
- On RAW deblurring with Stripformer on RealBlur-J, ParamISP reaches 30.70 dB PSNR and 0.8984 SSIM, above CycleISP (29.94 dB) and InvISP (28.82 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ParamNet with non-linear equalization and random optical parameter dropout (ablation) | ParamISP · D7000 sRGB→RAW | PSNR: 34.77 -> 36.21 | Helped |
| Add non-linear equalization to ParamNet (ablation) | ParamISP · D7000 sRGB→RAW | PSNR | Helped |
| Add random optical parameter dropout (ablation) | ParamISP · D7000 sRGB→RAW | PSNR: 35.64 -> 36.21 | Helped |
| Add sensor sensitivity (ablation) | ParamISP · D7000 sRGB→RAW | PSNR: 35.13 -> 36.21 | Helped |
| Add exposure time (ablation) | ParamISP · D7000 sRGB→RAW | PSNR: 35.25 -> 36.21 | Helped |
| Add aperture size (ablation) | ParamISP · D7000 sRGB→RAW | PSNR: 35.87 -> 36.21 | Helped |
| Add focal length (ablation) | ParamISP · D7000 sRGB→RAW | PSNR: 35.46 -> 36.21 | Helped |
| Replace CIE XYZ Net local mapping with LocalNet (ablation) | CIE XYZ Net · D7000 sRGB→RAW | PSNR: 30.04 -> 33.12 | Helped |
| Add GlobalNet (ablation) | CIE XYZ Net + LocalNet · D7000 sRGB→RAW | PSNR: 33.12 -> 33.66 | Helped |
| Use ParamISP instead of InvISP | sRGB-to-RAW reconstruction · D7000 | PSNR: 33.48 -> 38.49 | Helped |
| Use ParamISP instead of InvISP | RAW-to-sRGB reconstruction · D7000 | PSNR: 30.20 -> 34.14 | Helped |
| Use ParamISP instead of CycleISP | RAW deblurring · RealBlur-J test | PSNR: 29.94 -> 30.70 | Helped |
Surprising findings
- Adding ParamNet without non-linear equalization caused training to diverge, showing naive use of raw camera parameters is unstable.
Limitations
- Effect of aperture size and focal length on ISP behavior is not clearly understood.
- Pre-training across different cameras may introduce inconsistent ISP behavior that harms the pre-trained model.
- Each target camera requires separate training (camera-wise training).
Experimental setup
Datasets: RAISE RealBlur S7 ISP dataset Samsung Galaxy S22 captures
Benchmarks: RAISE S7 ISP dataset RealBlur-J
Metrics: PSNR SSIM
Baselines: UPI CIE XYZ Net CycleISP InvISP
Builds on: CIE XYZ Net CycleISP InvISP CBAM U-Net
Closest papers by meaning
- RGB-Event ISP: The Dataset and Benchmark (ICLR 2025 · similarity 0.84)
- RL-SeqISP: Reinforcement Learning-Based Sequential Optimization for Image Signal Processing (AAAI 2024 · similarity 0.83)
- ISPDiffuser: Learning RAW-to-sRGB Mappings with Texture-Aware Diffusion Models and Histogram-Guided Color Consistency (AAAI 2025 · similarity 0.83)
- RGB-Only Supervised Camera Parameter Optimization in Dynamic Scenes (NeurIPS 2025 · similarity 0.82)
- Learning Degradation-Independent Representations for Camera ISP Pipelines (CVPR 2024 · similarity 0.82)
- Leveraging Frame Affinity for sRGB-to-RAW Video De-rendering (CVPR 2024 · 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.