Research map / Detection and segmentation / CNN features
Dataset Enhancement with Instance-Level Augmentations
Orest Kupyn, Christian Rupprecht
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 with instance-level diffusion repainting augmentation (30% per instance) | Deformable-DETR · MS-COCO val | AP: 39.3 -> 40.5 | Improved |
| train with instance-level augmentation | YOLOv5m · MS-COCO val | AP: 44.1 -> 45.7 | Improved |
| train with instance augmentation in low-data regimes | YOLOv5m · MS-COCO (10% data) | AP: 25.8 -> 27.9 | Improved |
| pretrain on our synthetic VOC data then finetune on real VOC, vs real only | Mask2Former · Pascal VOC | mIoU: 77.3 -> 78.2 | Improved |
| our synthetic data instead of Diffumask synthetic data | Mask2Former · Pascal VOC (synthetic only) | mIoU: 57.4 -> 72.9 | Improved |
| train saliency models on our DUTS expansion | U2Net, F3-Net, TRACER-4 · ECSSD, DUTS-TE, DUT-OMRON, HKU-IS | F_max, MAE, S_m: 34/36 metrics improved | Improved |
| replace all people with synthetic people in COCO training | Deformable-DETR · COCO val (all classes) | AP: 39.3 -> 38.3 | Hurt |
| replace people with synthetic people | Deformable-DETR · COCO val (people only) | AP: 49.7 -> 45.6 | Hurt |
| anonymize people in DUTS | TRACER-0 · DUTS-TE | F_max: 0.889 -> 0.886 | No clear effect |
| add edge and depth ControlNet conditioning to inpainting (ablation) | instance augmentation pipeline · DUTS-TE | F_max: 0.886 -> 0.892 | Improved |
| add edge control on top of depth (ablation) | instance augmentation pipeline · DUTS-TE | F_max: 0.890 -> 0.892 | Improved |
| add prompt engineering (descriptions, colors, lighting) (ablation) | instance augmentation pipeline · DUTS-TE | F_max: 0.889 -> 0.892 | Improved |
Closest papers by meaning
- A Simple Background Augmentation Method for Object Detection with Diffusion Model (ECCV 2024 · similarity 0.91)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.86)
- Prompting Diffusion Models for Zero-Shot Instance Segmentation (NeurIPS 2026 · similarity 0.86)
- InstaGen: Enhancing Object Detection by Training on Synthetic Dataset (CVPR 2024 · similarity 0.85)
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.85)
- SNIDA: Unlocking Few-Shot Object Detection with Non-linear Semantic Decoupling Augmentation (CVPR 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.
