Research map / Probabilistic and scientific modeling / Variational and generative sampling
Towards Latent Attribute Discovery From Triplet Similarities
Ishan Nigam, Pavel Tokmakov, Deva Ramanan
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 end-to-end training of the embedding to MVTE (compare LSN with end-to-end MVTE) | LSN · Zappos-Human | Supervised-Eval accuracy: end-to-end MVTE 3.14% worse than LSN | Improved |
| compare LSN with end-to-end MVTE on Celeb-A | LSN · Celeb-A Faces | Supervised-Eval accuracy: end-to-end MVTE 4.72% worse than LSN | Improved |
| use hard assignment of triplets to latent embeddings instead of soft assignment (MVTE) | LSN · UT-Zappos Zappos-Human | Supervised-Eval accuracy: 77.94 -> 88.91 | Improved |
| learn multiple latent attribute embeddings instead of a single embedding (SSN) | LSN · Celeb-A Faces 8 attributes | Supervised-Eval accuracy: 71.40 -> 83.53 | Improved |
| add attribute supervision (CSN vs LSN) | LSN · Zappos-Human | Supervised-Eval accuracy: 88.91 -> 97.36 | Improved |
| increase number of underlying attributes from 2 to 8 | LSN · Celeb-A Faces | Supervised-Eval accuracy: 92.33 -> 83.53 | Hurt |
| use 4 latent embeddings instead of 2 on Zappos-Human (ablation) | LSN · Zappos-Human | Supervised-Eval accuracy | No clear effect |
| use 4 latent embeddings instead of 2 on Zappos-Meta (ablation) | LSN · Zappos-Meta | Supervised-Eval accuracy | Improved |
| replace learned attribute clustering with random assignment of 4 attributes to 2 embeddings (ablation) | LSN · Zappos-Human | Supervised-Eval accuracy: 6.69±0.53% lower with random assignment | Improved |
| replace learned attribute clustering with random assignment of 40 attributes to 8 embeddings (ablation) | LSN · Celeb-A Faces (40 attributes) | Supervised-Eval accuracy: 3.21 ± 0.68% lower with random assignment | Improved |
| increase latent embeddings beyond 8 on Celeb-A (40 attributes) (ablation) | LSN · Celeb-A Faces (40 attributes) | Supervised-Eval accuracy | Hurt |
Closest papers by meaning
- Effectively Leveraging Attributes for Visual Similarity (ICCV 2021 · similarity 0.88)
- Learning Similarity Conditions Without Explicit Supervision (ICCV 2019 · similarity 0.86)
- Conditional Similarity Networks (CVPR 2017 · similarity 0.86)
- Fine-Grained Fashion Similarity Learning by Attribute-Specific Embedding Network (AAAI 2020 · similarity 0.85)
- Intrinsic Dimension Correlation: uncovering nonlinear connections in multimodal representations (ICLR 2025 · similarity 0.83)
- Latent Zoning Network: A Unified Principle for Generative Modeling, Representation Learning, and Classification (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.
