Research map / Detection and segmentation / CNN features
Real-Time Scene-Adaptive Tone Mapping for High-Dynamic Range Object Detection
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 a neural photometric calibration plus scaling-invariant local tone mapper that converts 24-bit HDR RAW to LDR-like inputs for object detectors, trained end-to-end and transferred from sRGB, running at 4K real time on Jetson.
Problem
Object detectors trained on LDR sRGB images degrade severely or fail to converge on high-bit-depth HDR RAW automotive imagery, and traditional ISP and tone mapping are designed for human perception rather than machine vision.
Approach
A learnable neural photometric calibration normalizes HDR radiance into a unified scale, followed by a bias-free (scaling-invariant) CNN tone mapper that acts as a local tone mapper and is jointly optimized with the detector, with a performance-transfer finetuning scheme from LDR sRGB.
Key results
- On RoD with Faster R-CNN, the method reaches 49.8 mAP versus 46.1 for the best AI-ISP baseline (RAODNet) and 45.7 for HDR ISP, while raw HDR RAW input yields NaN.
- The Lite variant runs YOLOv3 at 4K on NVIDIA Jetson AGX Orin in 22 ms (45 FPS) with 16-bit float precision.
- Performance transfer with COCO-pretrained weights tunes only 33.6% of Faster R-CNN parameters and reaches 30.6 mAP versus 22.5 for AdaptiveISP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply Ours instead of RAODNet23 | Faster R-CNN on RoD · RoD | mAP: 46.1 -> 49.8 | Helped |
| Apply Ours instead of Mantiuk08 | YOLOv3 on RoD · RoD | mAP: 44.8 -> 45.1 | Helped |
| Add neural photometric calibration (ablation) | Faster R-CNN · RoD day | mAP: 36.7 -> 40.3 | Helped |
| Add neural photometric calibration (ablation) | Faster R-CNN · RoD night | mAP: 38.5 -> 45.8 | Helped |
| Add neural photometric calibration (ablation) | Faster R-CNN · RoD mixed scene | mAP: 40.7 -> 49.8 | Helped |
| Use a scaling-invariant tone mapper instead of a scaling-variant mapper (ablation) | Faster R-CNN · RoD to RhoVision | mAP: 14.3 -> 26.5 | Helped |
| Use NLPD instead of L1 for tone-mapper pretraining (ablation) | Faster R-CNN · RoD | mAP: 41.3 -> 49.8 | Helped |
| Vary the neural photometric calibration radiance range (ablation) | Faster R-CNN · RoD | mAP: 49.6 to 49.9 | No clear effect |
Surprising findings
- Directly feeding 24-bit HDR RAW images into detectors produces NaN training results, and the paper attributes this to gradient instability from extreme dynamic range.
Limitations
- Still requires a separate tone-mapping stage cascaded with the detector rather than a detector that consumes HDR directly.
- The HDR video dataset is not yet released; real-world validation is limited to a qualitative demo and a prototype.
- Evaluation is largely on a single HDR dataset (RoD) with a single sensor type.
Experimental setup
Datasets: RoD COCO Objects365 RhoVision
Benchmarks: RoD RhoVision
Metrics: mAP mAR AP50 AP75 latency (ms) FLOPs (G)
Baselines: Mantiuk08 CLAHE94 HDR ISP ReconfigISP21 Zero-DCE++ SCI AnscombeNet IANet22 RAODNet23 RawOrCooked23 RAWAdapter AdaptiveISP
Builds on: Faster R-CNN YOLOv3 Normalized Laplacian Pyramid Distance HDR ISP (HDRPlus) MMDetection
Closest papers by meaning
- Zero-Shot Structure-Preserving Diffusion Model for High Dynamic Range Tone Mapping (CVPR 2024 · similarity 0.88)
- Neural Exposure Fusion for High-Dynamic Range Object Detection (CVPR 2024 · similarity 0.88)
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.86)
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.86)
- Towards Real-World HDR Video Reconstruction: A Large-Scale Benchmark Dataset and A Two-Stage Alignment Network (CVPR 2024 · similarity 0.85)
- Color Event Enhanced Single-Exposure HDR Imaging (AAAI 2024 · similarity 0.85)
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.