Research map / Detection and segmentation / Open-set and few-shot detection
OrthoRF: Exploring Orthogonality in Object-Centric Representations
Despoina Touska, Bastiaan Onne Fagginger Auer, Alexandru Onose, Tejaswi Kasarla, Luis Armando Pérez Rey, Maximilian Lipp, Lyubov Amitonova, Martin R. Oswald, Pascal Cerfontaine
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 |
|---|---|---|---|
| add softmax with centering and orthogonality loss together to Rotating Features (ablation) | OrthoRF · 4Shapes (n=5) | ARI: 0.9750 -> 0.9995 | Improved |
| add orthogonality loss (lambda 0 -> 0.1) on top of softmax with centering (ablation) | OrthoRF · 4Shapes (n=5) | ARI: 0.6280 -> 0.9995 | Improved |
| add softmax with centering on top of orthogonality loss (ablation) | OrthoRF · 4Shapes (n=5) | ARI: 0.8530 -> 0.9995 | Improved |
| evaluate OrthoRF thresholded psi_final instead of k-means on z_final (ablation) | OrthoRF · 4Shapes (n=5) | MBO_OV_i: 0.8204 -> 0.9832 | Improved |
| OrthoRF vs RF under identical k-means post-processing on visible regions | OrthoRF · 4Shapes | ARI-BG, MBO_i | No clear effect |
| use large orientation dimension n=20 (ablation) | OrthoRF vs RF · 4Shapes | ARI-BG: RF 0.7129 vs OrthoRF 0.9974 | Improved |
| use n smaller than the number of objects (n=2,3) (ablation) | OrthoRF vs RF · 4Shapes | ARI-BG: OrthoRF 0.4987 vs RF 0.667 at n=2 | Hurt |
| orthogonality loss weight lambda 0 vs positive for small n (ablation) | OrthoRF · 4Shapes (n=5, n=6) | ARI: n=5: 0.6279 -> 0.9997 (lambda 0 -> 0.1) | Improved |
| orthogonality loss weight lambda at n=20 (ablation) | OrthoRF · 4Shapes (n=20) | ARI: lambda 0: 0.9975; lambda 0.1: 0.9988 | No clear effect |
| replace RF with OrthoRF on SEM data | OrthoRF · SEM noisy | ARI-BG: 0.7043 -> 0.8356 | Improved |
| replace RF with OrthoRF on Shapes dataset with variable object count | OrthoRF · Shapes | ARI-BG: 0.744 -> 0.833 | Improved |
| replace RF with OrthoRF on MNIST&Shape | OrthoRF · MNIST&Shape | ARI+BG: ARI-BG 0.972 -> 0.996; ARI+BG 0.699 -> 0.667 | Mixed |
Closest papers by meaning
- Bootstrapping Top-down Information for Self-modulating Slot Attention (NeurIPS 2024 · similarity 0.85)
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.84)
- A Data-Centric Revisit of Pre-Trained Vision Models for Robot Learning (CVPR 2025 · similarity 0.84)
- Orientation Matters: Making 3D Generative Models Orientation-Aligned (NeurIPS 2025 · similarity 0.84)
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery (NeurIPS 2026 · similarity 0.84)
- Temporally Consistent Object-Centric Learning by Contrasting Slots (CVPR 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.
