Visual Question Answering On Docvqa Test

評価指標

ANLS

評価結果

このベンチマークにおける各モデルのパフォーマンス結果

モデル名	ANLS	Paper Title	Repository
MatCha	0.742	MatCha: Enhancing Visual Language Pretraining with Math Reasoning and Chart Derendering
GPT-4	0.884	Layout and Task Aware Instruction Prompt for Zero-shot Document Image Question Answering
PaLI-3	0.876	PaLI-3 Vision Language Models: Smaller, Faster, Stronger
Qwen-VL	0.651	Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond
ERNIE-Layout large	0.8486	ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding
DUBLIN	0.782	DUBLIN -- Document Understanding By Language-Image Network	-
Pix2Struct-base	0.721	Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding
DUBLIN (variable resolution)	0.803	DUBLIN -- Document Understanding By Language-Image Network	-
PaLI-3 (w/ OCR)	0.886	PaLI-3 Vision Language Models: Smaller, Faster, Stronger
Qwen-VL-Plus	0.9024	Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond
PaLI-X (Single-task FT w/ OCR)	0.868	PaLI-X: On Scaling up a Multilingual Vision and Language Model
PaLI-X (Single-task FT)	0.80	PaLI-X: On Scaling up a Multilingual Vision and Language Model
Claude + LATIN-Prompt	0.8336	Layout and Task Aware Instruction Prompt for Zero-shot Document Image Question Answering
TILT-Large	0.8705	Going Full-TILT Boogie on Document Understanding with Text-Image-Layout Transformer
BERT_LARGE_SQUAD_DOCVQA_FINETUNED_Baseline	0.665	DocVQA: A Dataset for VQA on Document Images
DocFormerv2-large	0.8784	DocFormerv2: Local Features for Document Understanding
MLCD-Embodied-7B	0.916	Multi-label Cluster Discrimination for Visual Representation Learning
UDOP (aux)	0.878	Unifying Vision, Text, and Layout for Universal Document Processing
UDOP	0.847	Unifying Vision, Text, and Layout for Universal Document Processing
SMoLA-PaLI-X Generalist	0.906	Omni-SMoLA: Boosting Generalist Multimodal Models with Soft Mixture of Low-rank Experts	-

0 of 33 row(s) selected.