Research map / Detection and segmentation / Open-set and few-shot detection
GALA: Toward Geometry-and-Lighting-Aware Object Search for Compositing
Sijie Zhu, Zhe Lin, Scott Cohen, Jason Kuen, Zhifei Zhang, Chen Chen
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 foreground and background jointly instead of freezing the foreground network (ablation) | GALA foreground search · CAIS | mAP: 29.10 -> 17.93 | Hurt |
| add mask erosion/extension augmentation to direct training (ablation) | GALA foreground search · CAIS | mAP: 17.93 -> 23.45 | Improved |
| add alternating training of background and foreground networks (ablation) | GALA · CAIS | mAP: 29.99 -> 32.67 | Improved |
| add contrastive self-transformation loss (ablation) | GALA · CAIS | mAP: 31.20 -> 32.67 | Improved |
| use one round of alternating training instead of two rounds (ablation) | GALA alternating training · CAIS | mAP: 31.20 -> 30.35 | Hurt |
| train background then foreground instead of the reverse order (ablation) | GALA alternating training · CAIS | mAP: 28.82 -> 31.20 | Improved |
| add lighting transformation on top of geometry transformation (ablation) | GALA contrastive learning · CAIS | mAP: 32.11 -> 32.67 | Improved |
| use non-linear lighting instead of linear color jittering (ablation) | GALA contrastive learning · CAIS | mAP: 30.80 -> 32.67 | Improved |
| add geometry transformation contrast to alternating training (ablation) | GALA · CAIS | mAP: 31.20 -> 32.11 | Improved |
| switch from fixed-foreground baseline to full GALA, measured on geometry transformation ranking (ablation) | GALA · Pixabay | Recall@5 of original among transformed: 58.10 -> 98.55 | Improved |
| switch from fixed-foreground baseline to full GALA, measured on lighting transformation ranking (ablation) | GALA · Pixabay | Recall@5 of original among transformed: 53.40 -> 60.55 | Improved |
| replace UFO with GALA | unconstrained foreground search · CAIS | mAP-100: 32.17 -> 37.50 | Improved |
Closest papers by meaning
- Compositing-aware Image Search (ECCV 2018 · similarity 0.90)
- Unconstrained Foreground Object Search (ICCV 2019 · similarity 0.88)
- MV-CoLight: Efficient Object Compositing with Consistent Lighting and Shadow Generation (NeurIPS 2025 · similarity 0.85)
- ORIDa: Object-centric Real-world Image Composition Dataset (CVPR 2025 · similarity 0.83)
- OmniGlue: Generalizable Feature Matching with Foundation Model Guidance (CVPR 2024 · similarity 0.83)
- MetaGS: A Meta-Learned Gaussian-Phong Model for Out-of-Distribution 3D Scene Relighting (NeurIPS 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.
