Research map / Vision-language and multimodal / Human and embodied modeling
Separating Self-Expression and Visual Content in Hashtag Supervision
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 user-conditioned three-way tensor scoring to hashtag prediction | Image tagging models on YFCC100M · YFCC100M test set | Accuracy@1: 29.24% -> 41.24% | Helped |
| Add user-conditioned three-way tensor scoring to hashtag prediction | Image tagging models on YFCC100M · YFCC100M test set | Accuracy@10: 56.47% -> 70.75% | Helped |
| Add user-conditioned three-way tensor scoring instead of user-specific frequency counts | Image tagging on YFCC100M · YFCC100M test set | Accuracy@1: 38.07% -> 43.65% | Helped |
| Use MCLL rather than NCSL-MS for the user-agnostic image model | Hashtag-based image retrieval · YFCC100M test set | Precision@10: 6.0% -> 9.9% | Helped |
| Use multi-sense rather than PPMI hashtag embeddings with NCSL | Hashtag-based image retrieval · YFCC100M test set | Precision@10: 3.5% -> 6.0% | Helped |
| Train the vision network from scratch for hashtag prediction rather than use ImageNet features | User-agnostic image tagging · YFCC100M test set | Accuracy@1: 17.21% -> 29.24% | Helped |
| Use MCLL rather than NCSL-PPMI for user-agnostic image tagging | Hashtag prediction · YFCC100M test set | Accuracy@10: 47.70% -> 56.47% | Helped |
| Use a three-way tensor rather than an MLP to score user-specific image-tag predictions | User-specific hashtag tagging with NCSL-MS · YFCC100M test set | Accuracy@1: 41.66% -> 43.65% | Helped |
| Use a three-way tensor rather than an MLP to score user-specific image-tag predictions | User-specific hashtag tagging with NCSL-PPMI · YFCC100M test set | Accuracy@1: 37.31% -> 40.43% | Helped |
| Use a three-way tensor rather than an MLP to score user-specific image-tag predictions | User-specific hashtag tagging with MCLL · YFCC100M test set | Accuracy@1: 35.58% -> 41.24% | Helped |
| Increase user embedding dimensionality (ablation) | Tensor (NCSL-MS) · YFCC100M test set | Accuracy@1 | Helped |
| Use user-specific hashtag modeling rather than user-agnostic modeling | Image tagging on YFCC100M · Users with more than approximately 27 uploaded photos | Accuracy@1 | Helped |
Closest papers by meaning
- Learning User Representations for Open Vocabulary Image Hashtag Prediction (CVPR 2020 · similarity 0.90)
- Diverse Image Annotation (CVPR 2017 · similarity 0.88)
- Revisiting Weakly Supervised Pre-Training of Visual Perception Models (CVPR 2022 · similarity 0.87)
- Tagging Like Humans: Diverse and Distinct Image Annotation (CVPR 2018 · similarity 0.86)
- MagicLens: Self-Supervised Image Retrieval with Open-Ended Instructions (ICML 2024 · similarity 0.83)
- Tag2Text: Guiding Vision-Language Model via Image Tagging (ICLR 2024 · 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.