Research map / Detection and segmentation / CNN features
Neural Exposure Fusion 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.
Replaces image-space HDR exposure fusion with a learned, detection-driven feature-domain fusion via a local cross-attention module, trained end-to-end with a differentiable ISP and Faster-RCNN for automotive HDR object detection.
Problem
Conventional HDR pipelines fuse LDR exposures in image space for human perception, discarding information useful for downstream object detection in high dynamic range outdoor scenes.
Approach
A local cross-attention fusion block that, at each spatial location, attends across feature maps from multiple exposures using learned queries and a two-axis softmax, trained end-to-end with detection losses through a differentiable ISP and exposure control.
Key results
- Local cross-attention fusion achieves 44.2 mAP, outperforming Raw HDR (41.5) and Deep HDR (41.9) by 2.7 and 2.3 points.
- Local cross-attention fusion runs at 57 FPS (17.5 ms) on Jetson AGX Orin with 3 exposures, versus 16 FPS for Deep HDR.
- Ablations show 2-dim softmax (+1.0%) and normalized queries (+1.9%) improve over conventional attention design.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace image-space Raw HDR fusion with local cross-attention feature fusion | Faster R-CNN HDR object detection · Automotive HDR test set | mAP: 41.5 -> 44.2 | Helped |
| replace image-space Deep HDR fusion with local cross-attention feature fusion | Faster R-CNN HDR object detection · Automotive HDR test set | mAP: 41.9 -> 44.2 | Helped |
| replace local cross-attention fusion with max pooling (ablation) | Neural Exposure Fusion · Automotive HDR test set | mAP: 44.2 -> 43.2 | Hurt |
| replace local cross-attention fusion with 1 x 1 convolution (ablation) | Neural Exposure Fusion · Automotive HDR test set | mAP: 44.2 -> 38.5 | Hurt |
| replace local cross-attention fusion with 3 x 3 convolution (ablation) | Neural Exposure Fusion · Automotive HDR test set | mAP: 44.2 -> 37.9 | Hurt |
| use late fusion of detections rather than feature fusion (ablation) | Neural Exposure Fusion · Automotive HDR test set | mAP: 44.2 -> 43.0 | Hurt |
| use 2D softmax instead of 1D softmax in local cross-attention (ablation) | Local Cross Attention · Automotive HDR test set | mAP: +1.0% | Helped |
| normalize the learned queries (ablation) | Local Cross Attention · Automotive HDR test set | mAP: +1.9% | Helped |
| learn a projection matrix for the keys W K (ablation) | Local Cross Attention · Automotive HDR test set | mAP: +0.1% | Helped |
| use 4 exposures instead of 3 (ablation) | Local Cross Attention · Automotive HDR test set | mAP: 43.7 -> 44.4 | Helped |
| fuse exposure features at Conv1 instead of Conv4 (ablation) | Local Cross Attention · Automotive HDR test set | mAP: 44.2 -> 43.3 | Hurt |
| fuse exposure features at Conv2 instead of Conv4 (ablation) | Local Cross Attention · Automotive HDR test set | mAP: 44.2 -> 43.9 | Hurt |
Surprising findings
- Fusing at later backbone stages (Conv4) gives the best accuracy, but Conv1 fusion retains 43.3 mAP while running at 123 FPS, showing a runtime-accuracy tradeoff.
Limitations
- Evaluated only on a proprietary automotive HDR dataset from one sensor and test vehicle.
- Runs on a general-purpose GPU whereas ISPs typically run on specialized ASICs, so runtime comparison is not like-for-like.
- Fusion gains are modest (about 2-3 mAP) and fusing earlier layers trades accuracy for speed.
Experimental setup
Datasets: automotive HDR images captured with Sony IMX490 sensor (18790 training examples) 1996-pair HDR test set with manual 2D bounding box annotations
Benchmarks: in-house automotive HDR test set (1996 frame pairs)
Metrics: AP at 50% IoU mAP
Baselines: Raw HDR Deep HDR Debevec and Malik exposure fusion Hanji et al. (PPNE) Shim et al. gradient auto-exposure Onzon et al. neural exposure control Max Pooling Fusion Convolutional fusion Late Fusion
Builds on: Faster R-CNN Transformer attention Neural exposure control (Onzon et al.) Deep HDR
Closest papers by meaning
- Real-Time Scene-Adaptive Tone Mapping for High-Dynamic Range Object Detection (NeurIPS 2025 · similarity 0.88)
- Self-Supervised High Dynamic Range Imaging with Multi-Exposure Images in Dynamic Scenes (ICLR 2024 · similarity 0.86)
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.86)
- Color Event Enhanced Single-Exposure HDR Imaging (AAAI 2024 · similarity 0.84)
- UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion (CVPR 2025 · similarity 0.84)
- SMR-Net: Semantic-Guided Mutually Reinforcing Network for Cross-Modal Image Fusion and Salient Object Detection (AAAI 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.