Research map / Image generation / GANs
Dark Side Augmentation: Generating Diverse Night Examples for Metric Learning
Albert Mohwald, Tomas Jenicek, Ondřej Chum
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 |
|---|---|---|---|
| train retrieval with GAN-generated night anchors instead of GeM baseline | GeM VGG-16 · Tokyo 24/7 | mAP: 79.4 -> 90.2 | Improved |
| train retrieval with GAN-generated night anchors instead of GeM baseline | GeM ResNet-101 · Tokyo 24/7 | mAP: 85.0 -> 92.2 | Improved |
| use synthetic GAN night images instead of real paired day-night data | GeM VGG-16 · Tokyo 24/7 | mAP: 87.0 -> 90.2 | Improved |
| add diverse anchor mining (ablation) | HEDN GAN C · Tokyo 24/7 | mAP: 88.7 -> 88.8 | Improved |
| add diverse anchor mining (ablation) | CLAHE C baseline · Tokyo 24/7 | mAP: 85.4 -> 85.9 | Improved |
| add CLAHE normalization (ablation) | HEDN GAN · Tokyo 24/7 | mAP: 88.0 -> 88.7 | Improved |
| train HED edge detector jointly (HEDN) instead of frozen HED (ablation) | EdgeMAC · Tokyo 24/7 | mAP: 79.5 -> 81.9 | Improved |
| add real night images (SfM-N/D) to GAN-generated training data (ablation) | CycleGAN C D · Tokyo 24/7 | mAP: 90.2 -> 88.6 | No clear effect |
| use 50% instead of 25% night anchors (ablation) | retrieval training · Tokyo 24/7 | mAP | No clear effect |
| replace GAN generator with simple LAB intensity inversion augmentation (ablation) | GeM · retrieval average over Tokyo, ROxf, RPar | avg mAP: 70.0 -> 71.2 (far below trained generators) | Mixed |
| translate night queries to day at inference with generator | retrieval at test time · Tokyo 24/7 | mAP: 39 and 52 mAP (substantially below baseline) | Hurt |
| apply color-invariant CIConv to ResNet-101 GeM | GeM ResNet-101 · Revisited Oxford | mAP: 65.3 -> 62.0 | Hurt |
Closest papers by meaning
- No Fear of the Dark: Image Retrieval Under Varying Illumination Conditions (ICCV 2019 · similarity 0.86)
- Unsupervised Monocular Depth Estimation for Night-time Images using Adversarial Domain Feature Adaptation (ECCV 2020 · similarity 0.85)
- DANNet: A One-Stage Domain Adaptation Network for Unsupervised Nighttime Semantic Segmentation (CVPR 2021 · similarity 0.85)
- DNIT: Enhancing Day-Night Image-to-Image Translation through Fine-Grained Feature Handling (Student Abstract) (AAAI 2024 · similarity 0.83)
- The Parables of the Mustard Seed and the Yeast: Extremely Low-Budget, High-Performance Nighttime Semantic Segmentation (AAAI 2025 · similarity 0.83)
- NightAdapter: Learning a Frequency Adapter for Generalizable Night-time Scene Segmentation (CVPR 2025 · similarity 0.82)
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.
