IDEA Lab

nonprofit

china · Founded 2020

The International Digital Economy Academy (IDEA) is a research institute in Shenzhen founded in November 2020 by Harry Shum (Shen Xiangyang), former Executive Vice President of Microsoft and a founding member of Microsoft Research Asia. The academy is backed by the Shenzhen municipal government as part of the city's push to become a global AI innovation hub.

IDEA operates two major research tracks. IDEA-Research (computer vision), led by Lei Zhang, has produced world-class object detection models including Grounding DINO, DINO-DETR, Grounded SAM, and DINO-X. These models are among the most widely adopted open-source detection frameworks globally, with Grounding DINO alone accumulating tens of thousands of GitHub stars.

The second track, IDEA-CCNL (NLP/cognitive computing), produced the Fengshenbang ecosystem and the Ziya series of Chinese language models. While the CV work has garnered wider international recognition, the NLP division has contributed important Chinese-language pretraining infrastructure and fine-tuning frameworks.

Website HuggingFace GitHub
visionopen-sourceobject-detectionnlpnonprofit

People

Outputs (18)

Rex-Omni 2

model paper

3B-parameter MLLM that redefines object detection and visual perception as next-token prediction. Unifies detection, OCR, pointing, keypointing, and v…

↓ 27.6k

DINO-X 2

model paper

Unified vision model for open-world object detection and understanding. Supports text, visual, and customized prompts including prompt-free universal…

📄 4

MotionCLR 2

library paper

Motion generation and training-free editing via understanding attention mechanisms. Self-attention finds word-sequence correspondence while cross-atte…

MotionLLM 2

model paper

Multi-modal human behavior understanding leveraging LLMs for both video and motion (SMPL) modalities. Introduces MoVid dataset and MoVid-Bench for hum…

📄 5

Grounding DINO 1.5 2

model paper

Advanced open-set object detection suite with Pro and Edge variants. Pro scales up architecture with enhanced vision backbone and 20M+ training images…

📄 10

T-Rex 3

model paper

Interactive object counting and detection via visual prompting. Users mark points or boxes on reference images, then T-Rex detects all similar objects…

📄 8

detrex 2

library paper

Open-source research platform for Transformer-based detection algorithms built on Detectron2. Provides unified, modular codebase supporting 15+ mainst…

📄 15

Ziya LLM 4

model

Chinese-English bilingual LLM series built on LLaMA. Three-stage training: large-scale pre-training, multi-task supervised fine-tuning, and RLHF. Supp…

↓ 1.2k

Stable-DINO 2

library paper

Detection Transformer with Stable Matching. Addresses multi-optimization path problem in DETR caused by one-to-one matching instability. Proposes posi…

📄 5

Grounded SAM 3

library paper

Marries Grounding DINO with Segment Anything for automated open-world detection, segmentation, and generation. Achieves 48.7 mean AP on SegInW zero-sh…

📄 88

OpenSeeD 2

library paper

A Simple Framework for Open-Vocabulary Segmentation and Detection. Jointly learns from different segmentation and detection datasets with a pre-traine…

📄 2

Grounding DINO 2

model paper

Open-set object detector that marries DINO with grounded pre-training. Detects arbitrary objects using text prompts (category names or referring expre…

↓ 1.4M 📄 240

Taiyi 3

model paper

First open-source Chinese Stable Diffusion model family. Trained on 20M filtered Chinese image-text pairs from Noah-Wukong and Zero datasets. Taiyi-Di…

↓ 510 📄 2

Mask DINO 2

library paper

Unified Transformer-based framework for object detection and segmentation. Extends DINO with a mask prediction branch supporting instance, panoptic, a…

📄 19

DINO (DETR) 2

library paper

DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. Uses contrastive denoising training, look-forward-twice box prediction, and…

📄 747

DN-DETR 2

library paper

Accelerates DETR training by introducing query denoising. Feeds noisy ground-truth boxes into the decoder and trains the model to reconstruct original…

📄 54

DAB-DETR 2

library paper

Dynamic Anchor Boxes are Better Queries for DETR. Proposes using dynamic anchor boxes as positional queries in DETR, making the queries more interpret…

📄 391

Fengshenbang 3

library model

Open-source large model ecosystem from IDEA-CCNL covering NLU, NLG, and cross-modal tasks. Includes three model families: Erlangshen (NLU, largest pub…

↓ 339