Research map / Detection and segmentation / Open-set and few-shot detection
Improving Viewpoint-Independent Object-Centric Representations through Active Viewpoint Selection
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 AVS, a multi-viewpoint object-centric model built on Latent Slot Diffusion that actively selects the next observation viewpoint by predicting unknown views and picking the one whose slots change most.
Problem
Multi-viewpoint object-centric learning typically uses random or sequential viewpoint selection, which may be inefficient or miss essential scene details; existing diffusion-based object-centric decoders are limited to single images or videos and cannot synthesize novel views.
Approach
An active viewpoint selection strategy that uses a pretrained slot-conditioned diffusion decoder to predict images from unknown viewpoints, then selects the candidate whose updated slots have the lowest cosine similarity to the previous slots (largest information gain) as the next observation; slots are made viewpoint-independent by averaging across viewpoints in multi-viewpoint slot attention.
Key results
- On CLEVRTEX, Ours (Active) reaches ARI-O 86.1, mIoU 54.3 and FID 80.3, versus 84.9/52.7/89.1 for random selection and 78.1/45.7/110.9 for LSD.
- On GSO, Ours (Active) achieves ARI-A 68.9 and mIoU 64.4, versus 65.3 and 62.3 for random selection.
- On ShapeNet, Ours (Active) improves ARI-O to 75.2 from 71.6 for random selection, with FID 99.2 versus 103.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use active rather than random viewpoint selection (ablation) | AVS · CLEVRTEX | ARI-A: 23.9 -> 24.3 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · CLEVRTEX | ARI-O: 84.9 -> 86.1 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · CLEVRTEX | mIoU: 52.7 -> 54.3 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · CLEVRTEX | LPIPS: 0.178 -> 0.175 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · CLEVRTEX | FID: 89.1 -> 80.3 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · GSO | ARI-A: 65.3 -> 68.9 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · GSO | ARI-O: 79.6 -> 82.2 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · GSO | mIoU: 62.3 -> 64.4 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · GSO | LPIPS: 0.176 -> 0.172 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · GSO | FID: 96.5 -> 87.2 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · ShapeNet | ARI-A: 54.7 -> 58.0 | Helped |
| use active rather than random viewpoint selection (ablation) | AVS · ShapeNet | ARI-O: 71.6 -> 75.2 | Helped |
Surprising findings
- On CLEVRTEX, ARI-A is lower for the proposed method (24.3) than for OCLOC (56.9) and LSD (52.9), which the authors attribute to background being split into two parts.
- LSD achieves slightly better LPIPS than the proposed method on all datasets, which the authors attribute to viewpoint-specific slots.
Limitations
- Active selection has high training complexity, scaling with the number of selected viewpoints and diffusion sampling steps.
- Training time (4.5 days on 4 RTX 4090) is longer than the baselines.
- On CLEVRTEX the ARI-A score is low due to background splitting from complex lighting and textures.
Experimental setup
Datasets: CLEVRTEX GSO ShapeNet
Benchmarks: CLEVRTEX GSO ShapeNet
Metrics: ARI-A ARI-O mIoU LPIPS FID
Baselines: SIMONe OCLOC LSD Ours (Random)
Builds on: Latent Slot Diffusion Slot Attention DINOSAUR OCLOC Stable Diffusion
Closest papers by meaning
- Learning Object-Centric Representations of Multi-Object Scenes from Multiple Views (NeurIPS 2020 · similarity 0.87)
- PEERING INTO THE UNKNOWN: ACTIVE VIEW SELECTION WITH NEURAL UNCERTAINTY MAPS FOR 3D RECONSTRUCTION (ICLR 2026 · similarity 0.87)
- Coverage Optimization for Camera View Selection (CVPR 2026 · similarity 0.86)
- Shepherding Slots to Objects: Towards Stable and Robust Object-Centric Learning (CVPR 2023 · similarity 0.86)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.84)
- WildCAT3D: Appearance-Aware Multi-View Diffusion in the Wild (NeurIPS 2025 · similarity 0.83)
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.