YOLO-WORLD: Custom Object Detection Without Retraining

In this video, I will explain how Yolo-World works. A Real-Time Open-Vocabulary Object Detection method for detecting new objects without the need to retrain the model.
YOLO-World is pre-trained on large-scale vision-language datasets such as Objects365, GQA, Flickr30K, and CC3M, providing it with strong zero-shot open-vocabulary capability and image grounding ability.
YOLO-World achieves fast inference speeds, and we demonstrate re-parameterization techniques for faster inference and deployment based on user vocabularies.
Github: github.com/AILab-CVC/YOLO-World
00:48 - YOLO-WORLD Paper
04:47 - YOLO-WORLD Architecture
08:47 - Some Examples
10:27 - HuggingFace Demo
12:03 - Google Colab Demo

[@rickyS-D76]

Nice, can you please make a video on detecting objects from a video using yolo world with boudiing box and confidence score. Thanks

[@nitishreddynandyala2609]

very useful, thank you

[@fintech1378]

is it vision language model like CLIP?

[@next_phase]

yes, it is similar to that. But this also provides bounding boxes.

[@informative7410]

How to convert it into tflite ??

[@next_phase]

I could not figure out a way to do that. I tried to do it using ultralytics but did not work out. You can check my code here: colab.research.google.com/drive/1cxqMiSuNPSwy4z8nhgxIVK8YQxINbak_?usp=sharing

[@user-rg7mm7bh5t]

The accuracy of yolo-world is too low.

[@next_phase]

I agree. However, don’t forget that this is the first version of YOLO with such ability. The next versions will be better.