Research map / Vision-language and multimodal / Human and embodied modeling
A Generative Approach for Wikipedia-Scale Visual Entity Recognition
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 Generative Entity Recognition (GER), which maps an image to a short discriminative language-based code for one of 6M Wikipedia entities, using GIT-based auto-regressive decoding with entity-based pretraining on WebLI.
Problem
Web-scale visual entity recognition: mapping a query image to one of ~6 million Wikipedia entities (OVEN benchmark), where classifiers have huge parameter counts and dual encoders lose fine-grained information and need dense storage.
Approach
Represent each entity as a short integer code built from the least-frequent, most discriminative tokens of its tokenized name (ALD codes), and train an auto-regressive image-to-code decoder; pretrain by retrieving WebLI images for each Wikipedia entity and replacing captions with entity names.
Key results
- GER-ALD reaches 22.7 HM on OVEN test after finetuning, versus 16.0 for PaLI-17B, using about 42x fewer parameters (0.40B).
- GER-ALD outperforms GER-ATOMIC by 14% relative with 55M pretraining images and by over 1000% when pretraining data is reduced 5x.
- On ImageNet-LT and WebVision, GER-ALD and GER-ATOMIC reach 80.9 and 84.8 accuracy, state of the art without extra data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GER-ALD instead of PaLI-17B | OVEN entity recognition · OVEN test | HM top-1 accuracy: +6.8 top-1 HM test accuracy | Helped |
| Use unambiguous language-based discriminative codes rather than atomic codes (ablation) | GER · OVEN test | HM top-1 accuracy | Helped |
| Use unambiguous language-based discriminative codes rather than tokenized entity names (ablation) | GER · OVEN test | HM top-1 accuracy | Helped |
| Use entity-based WebLI pretraining rather than standard WebLI caption pretraining (ablation) | GER · OVEN validation | top-1 accuracy: +11.1 | Helped |
| Replace retrieved WebLI captions with corresponding OVEN entity names during pretraining (ablation) | GER · OVEN validation | top-1 accuracy: +1.9 | Helped |
| Use GER-ALD codes rather than tokenized entity names for pretraining (ablation) | GER · OVEN validation | top-1 accuracy: +2.7 | Helped |
| Select least frequent entity-name tokens for codes rather than most frequent, first, or random tokens (ablation) | GER-ALD · OVEN validation | HM top-1 accuracy | Helped |
| Decode selected GER-ALD tokens least-frequent-first rather than least-frequent-last (ablation) | GER-ALD · OVEN validation | HM top-1 accuracy | Helped |
| Increase GER-ALD code length beyond 4 (ablation) | GER-ALD · OVEN validation | top-1 accuracy | Hurt |
| Use semantic codes rather than atomic codes with a million-scale label space (ablation) | GER · OVEN validation | relative improvement: +26% relative improvement | Helped |
| Reduce the amount of pretraining data for atomic codes (ablation) | GER-ATOMIC · OVEN validation | relative improvement of GER-ALD over GER-ATOMIC: more than 1000% | Mixed |
| Use GER-ATOMIC rather than a standard classification MLP (ablation) | ImageNet-LT classification · ImageNet-LT | accuracy: 74.3 versus 81.0 | Helped |
Surprising findings
- Atomic (unstructured) codes fail badly when pretraining data or model capacity is reduced or when the label space reaches the million scale.
- Decoding the full tokenized entity name performs slightly worse than decoding only the discriminative least-frequent tokens, even though the same information is present.
Limitations
- Entity-WebLI pretraining dataset is not released and relies on retrieval from WebLI and LAION.
- Codes are fixed heuristically from name tokens rather than learned; GER-HKC relies on an off-the-shelf text encoder.
- Ablations are run on the validation set after pretraining only, and some gains depend on the pretraining data scale.
Experimental setup
Datasets: OVEN WebLI LAION ImageNet-LT WebVision
Benchmarks: OVEN ImageNet-LT WebVision
Metrics: top-1 accuracy harmonic mean of seen and unseen top-1 accuracy
Baselines: CLIP-L/14 dual encoder CLIP2CLIP GiT-Large PaLI-3B PaLI-17B GER-ATOMIC GER-CAPTION GER-HKC hierarchical classifier visual matching
Builds on: GIT DSI OVEN CLIP
Closest papers by meaning
- Web-Scale Visual Entity Recognition: An LLM-Driven Data Approach (NeurIPS 2024 · similarity 0.90)
- A Dual-Way Enhanced Framework from Text Matching Point of View for Multimodal Entity Linking (AAAI 2024 · similarity 0.86)
- EvdCLIP: Improving Vision-Language Retrieval with Entity Visual Descriptions from Large Language Models (AAAI 2025 · similarity 0.85)
- Reverse Region-to-Entity Annotation for Pixel-Level Visual Entity Linking (AAAI 2025 · similarity 0.85)
- Multi-Grained Query-Guided Set Prediction Network for Grounded Multimodal Named Entity Recognition (AAAI 2025 · similarity 0.85)
- MOFI: Learning Image Representations from Noisy Entity Annotated Images (ICLR 2024 · 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.