Research map / Detection and segmentation / Open-set and few-shot detection
BOOTPLACE: Bootstrapped Object Placement with Detection Transformers
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.
Proposes BOOTPLACE, which reframes copy-paste object placement as placement-by-detection using a detection transformer on object-subtracted backgrounds plus an object-to-region association network, trained with bootstrapped random object subtraction.
Problem
Learning precise object placement for copy-paste image composition, where dense ground-truth location supervision is sparse and prior generative or relaxed transformer methods give imprecise or implausible placements.
Approach
Treats object placement as detection of regions of interest on object-subtracted backgrounds (with scene object locations encoded as input), then associates each compositing object query to a detected region via a differentiable negative-correlation association score, and augments data by randomly subtracting and recomposing objects.
Key results
- On OPA, top-5 IOU reaches 0.281 versus 0.241 for TopNet and top-5 IOU50 of 22.41 versus 15.95.
- On Cityscapes, top-5 IOU50 reaches 6.09 versus 1.61 for TopNet, with top-5 IOU 0.281 versus 0.070.
- Ablation shows explicit location encoder raises IOU@5 by 0.195 and single-label supervision lowers it by 0.042 relative to full multi-object supervision in the text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Gaussian smoothing (ablation) | BOOTPLACE · Cityscapes | IOU@5 (%): 0.082 -> 0.281 | Helped |
| add bootstrapped data augmentation (ablation) | BOOTPLACE · Cityscapes | IOU@5 (%): 0.121 -> 0.281 | Helped |
| use negative rather than positive correlations (ablation) | BOOTPLACE association network · Cityscapes | IOU@5 (%): 0.125 -> 0.281 | Helped |
| use multi-object rather than single-object supervision (ablation) | BOOTPLACE · Cityscapes | IOU@5 (%): 0.239 -> 0.281 | Helped |
| add the location encoder (ablation) | BOOTPLACE · Cityscapes | IOU@5 (%): 0.086 -> 0.281 | Helped |
| use BOOTPLACE instead of TopNet | object repositioning · Cityscapes | IOU@5 (%): 0.070 -> 0.281 | Helped |
| use BOOTPLACE instead of GracoNet | object repositioning · OPA | IOU50@5 (%): 16.60 -> 22.41 | Helped |
| use BOOTPLACE instead of TopNet | object repositioning · OPA | IOU@1 (%): 0.197 -> 0.197 | No clear effect |
Surprising findings
- Authors report IOU50 is largely unchanged or lower for some ablations (e.g., 0 for removing smoothing at top-1), suggesting metric sensitivity on small-scale results.
Limitations
- Detects locations in parallel, so it is unsuited to sequential placement and can cause occlusions or collisions.
- Lacks out-of-plane rotation and perspective modeling, giving less realistic compositions.
- Only single-object labels used on OPA due to lack of multi-object annotations.
Experimental setup
Datasets: Cityscapes OPA Mapillary Vistas
Benchmarks: Cityscapes OPA Mapillary Vistas
Metrics: IOU50 top-1 IOU top-5 IOU user study plausibility bbox scale/x-center/y-center diversity std
Baselines: PlaceNet GracoNet SAC-GAN TopNet
Builds on: DETR TopNet Faster R-CNN
Closest papers by meaning
- TopNet: Transformer-Based Object Placement Network for Image Compositing (CVPR 2023 · similarity 0.88)
- Proposal-Contrastive Pretraining for Object Detection from Fewer Data (ICLR 2023 · similarity 0.85)
- MonoPlace3D: Learning 3D-Aware Object Placement for 3D Monocular Detection (CVPR 2025 · similarity 0.84)
- ODGEN: Domain-specific Object Detection Data Generation with Diffusion Models (NeurIPS 2024 · similarity 0.84)
- DiffuBox: Refining 3D Object Detection with Point Diffusion (NeurIPS 2024 · similarity 0.84)
- Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · similarity 0.84)
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.