Research map / Detection and segmentation / Open-set and few-shot detection

ORIDa: Object-centric Real-world Image Composition Dataset

CVPR 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Object-centric learning theory and video · 2 citations (OpenAlex)

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.

Introduces ORIDa, a publicly released real-captured dataset of over 30,000 images of 200 objects in factual-counterfactual sets and factual-only scenes, and fine-tunes Stable Diffusion inpainting on it for object removal and insertion.

Problem

Object compositing needs large, diverse, real-world training and evaluation data; existing datasets are small, benchmark-only, synthetic, or not publicly available.

Approach

A large-scale real-captured dataset with factual-counterfactual sets (four object positions plus a background-only image per scene) and factual-only scenes, with raw DNG files enabling ISP augmentation; models trained on it without synthetic data.

Key results

• On the ORIDa held-out test set, SD-Ours-r object removal reaches 25.60 PSNR and 0.088 LPIPS versus 21.76 PSNR and 0.108 LPIPS for SD-Inpaint.
• In a 76-participant user study for object removal, SD-Ours-r scored 4.23/5 versus 2.78 for SD-Inpaint, 2.63 for LaMa, and 1.96 for MGIE.
• In a 62-participant preference study for object insertion, SD-Ours-i was preferred 66% for identity, 79% for shadow generation, 71% for color harmonization, and 67% for overall quality.

Measured effects reported in the paper

Limitations

• Object poses were limited during capture, reducing pose variation.
• Scene-level generalization evaluated only with a single fine-tuned SD-Inpaint model and a limited set of baselines.

Experimental setup

Datasets: ORIDa COCO

Benchmarks: ORIDa held-out test set COCO images (qualitative and user study)

Metrics: PSNR DINO similarity CLIP similarity LPIPS user study rating user preference rate

Baselines: SD-Inpaint LaMa MGIE Paint-by-Example AnyDoor ObjectStitch Copy & Paste

Builds on: ObjectDrop Stable Diffusion inpainting

Closest papers by meaning

• Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · similarity 0.84)
• OverLay++: Dense-Overlap Layout-to-Image Generation Dataset (NeurIPS 2026 · similarity 0.84)
• Orientation Matters: Making 3D Generative Models Orientation-Aligned (NeurIPS 2025 · similarity 0.84)
• ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.84)
• unMORE: Unsupervised Multi-Object Segmentation via Center-Boundary Reasoning (ICML 2025 · similarity 0.83)
• The Scene Language: Representing Scenes with Programs, Words, and Embeddings (CVPR 2025 · similarity 0.83)