Lei Zhang

Chief Scientist, IDEA Research. Formerly @Microsoft, @MSFTResearch

[Lei Zhang @leizhangcs]

Referring is a critical skill for machines to perform language-guided tasks. I'm excited to share our latest advancement, #DINOXSeek, which was built on our new research work #RexSeek!

Paper: https://t.co/bngmsuqJp9
Github: https://t.co/vG4pv9BT2L Demo: https://t.co/BbQjPvlWyF

[Lei Zhang @leizhangcs]

The API for prompt-free object detection is ready, which means you do not need to provide any prompt and DINO-X will automatically recognize, detect, and segment objects in the provided images. Feel free to check out https://t.co/4V7x4tiZax and try this feature.

[Lei Zhang @leizhangcs]

🌟 Introducing DINO-X, our groundbreaking unified vision model at IDEA Research!

Paper: https://t.co/89gUTzuRVV
Blog: https://t.co/zUPG5J4kYx
Playground: https://t.co/KHBfWOyG5P
API: https://t.co/35hbTnR8mK

#AI #ComputerVision #ObjectDetection #IDEAResearch

Highlights:

[Lei Zhang @leizhangcs]

@DanielXieee Not ready yet. It's in our roadmap.

[Lei Zhang @leizhangcs]

🌟 Introducing DINO-X, our groundbreaking unified vision model at IDEA Research!

Paper: https://t.co/89gUTzuRVV
Blog: https://t.co/zUPG5J4kYx
Playground: https://t.co/KHBfWOyG5P
API: https://t.co/35hbTnR8mK

#AI #ComputerVision #ObjectDetection #IDEAResearch

Highlights:

[Lei Zhang @leizhangcs]

@Art_Intelligo It's so embarrassing. The model is able to detect the "goal" if given a prompt "goal", but it's prompt-free mode misses to suggest there is a "goal" in the picture. There is still space for future improvement. :)

[Lei Zhang @leizhangcs]

By adding a small language model on top of the object detection results, we can now describe each object with dense captions.

[Lei Zhang @leizhangcs]

In DINO-X, we add a new feature of prompt-free detection, making it possible to detect anything without the need of providing any prompt.

[Lei Zhang @leizhangcs]

✅ Best-in-class open-world detection performance
✅ Prompt-free detection of anything
✅ Powered by 100M+ grounding training images
✅ Multi-task support: detection, segmentation, pose estimation, object captioning & more

See more examples

[Visincept (Formly IDEA-CVR) @visincept]

⚠️T-rex Label v1.2 Update is here⚠️
✅You can now manually get rid of extra boxes in smart drawing
✅Shortcut keys have been improved
✅Creating categories is faster and easier

more about product? :: https://t.co/968qhdQMpZ

#computersciences #CV #annotation #TREX

[Shilong Liu @atasteoff]

Cool work!

Enjoy the new Grounded SAM 2

[Hongyang Li (弘洋) @ARChaser_]

#ECCV2024
Beyond Semantic, Tracking Any Area.

Tracking of the finest visual granularity greatly enhances the flexibility, allowing us to beyond the semantic and track any area.
Here, we use this demo as an example to demonstrate the potential of TAPTR in downstream tasks.

[Lei Zhang @leizhangcs]

We have a booth #2018 at CVPR this year. Come and visit us if you are interested in our new models Grounding DINO 1.5 and T-Rex2. #CVPR2024

[Visincept (Formly IDEA-CVR) @visincept]

We will be participating in the #CVPR2024 tomorrow! Join us at Booth #2018 to explore our latest models and products. Chat with the core Grounding DINO team members!
https://t.co/v3jZEJJnzW
@CVPR @CVPRConf

[Lei Zhang @leizhangcs]

The model can be tried out on the playground and accessed by open API.

Paper: https://t.co/XDJ9hv5BYB
Blog: https://t.co/P3C1r8T9UI
Playground: https://t.co/pjpHK09rw1
API Access: https://t.co/YRswXi8pY7

[Lei Zhang @leizhangcs]

Today, we're thrilled to announce the release of Grounding DINO 1.5, featuring two distinct models tailored for various scenarios:

🌟 Grounding DINO 1.5 Pro offers top-tier zero-shot performance on diverse benchmarks: 54.3 AP on COCO, 55.7 AP on LVIS, and 30.2 AP on ODinW.

[Lei Zhang @leizhangcs]

🌟 Grounding DINO 1.5 Edge is optimized for edge computing scenarios, achieving 75.2 FPS on NVIDIA A100 and over 10 FPS on NVIDIA Orin NX while maintaining an excellent balance between efficiency and accuracy.

[Lei Zhang @leizhangcs]

@DrawingDeskApp #MadeWithDrawingDeskl。dlr， https://t.co/bbGnN9AdXn