Tianheng Cheng^2,3,*, Lin Song^1,📧,*, Yixiao Ge^1,🌟,2, Wenyu Liu³, Xinggang Wang^3,📧, Ying Shan^1,2

* Equal contribution 🌟 Project lead 📧 Corresponding author

¹ Tencent AI Lab, ² ARC Lab, Tencent PCG ³ Huazhong University of Science and Technology

Updates

🔥[2024-2-3]: We support the Gradio demo now in the repo and you can build the YOLO-World demo on your own device!
[2024-2-1]: We've released the code and weights of YOLO-World now!
[2024-2-1]: We deploy the YOLO-World demo on HuggingFace 🤗, you can try it now!
[2024-1-31]: We are excited to launch YOLO-World, a cutting-edge real-time open-vocabulary object detector.

TODO

YOLO-World is under active development and please stay tuned ☕️!

• Gradio demo!
• Complete documents for pre-training YOLO-World.
• COCO & LVIS fine-tuning.
• Extra pre-trained models on more data, such as CC3M.
• Deployment toolkits, e.g., ONNX or TensorRT.
• Inference acceleration and scripts for speed evaluation.
• Automatic labeling framework for image-text pairs, such as CC3M.

Highlights

This repo contains the PyTorch implementation, pre-trained weights, and pre-training/fine-tuning code for YOLO-World.

• YOLO-World is pre-trained on large-scale datasets, including detection, grounding, and image-text datasets.

• YOLO-World is the next-generation YOLO detector, with a strong open-vocabulary detection capability and grounding ability.

• YOLO-World presents a prompt-then-detect paradigm for efficient user-vocabulary inference, which re-parameterizes vocabulary embeddings as parameters into the model and achieve superior inference speed. You can try to export your own detection model without extra training or fine-tuning in our online demo!

Abstract

The You Only Look Once (YOLO) series of detectors have established themselves as efficient and practical tools. However, their reliance on predefined and trained object categories limits their applicability in open scenarios. Addressing this limitation, we introduce YOLO-World, an innovative approach that enhances YOLO with open-vocabulary detection capabilities through vision-language modeling and pre-training on large-scale datasets. Specifically, we propose a new Re-parameterizable Vision-Language Path Aggregation Network (RepVL-PAN) and region-text contrastive loss to facilitate the interaction between visual and linguistic information. Our method excels in detecting a wide range of objects in a zero-shot manner with high efficiency. On the challenging LVIS dataset, YOLO-World achieves 35.4 AP with 52.0 FPS on V100, which outperforms many state-of-the-art methods in terms of both accuracy and speed. Furthermore, the fine-tuned YOLO-World achieves remarkable performance on several downstream tasks, including object detection and open-vocabulary instance segmentation.

Main Results

We've pre-trained YOLO-World-S/M/L from scratch and evaluate on the LVIS val-1.0 and LVIS minival. We provide the pre-trained model weights and training logs for applications/research or re-producing the results.

Zero-shot Inference on LVIS dataset

model	Pre-train Data	APfixed	APmini	APr	APc	APf	APval	APr	APc	APf	weights
YOLO-World-S	O365+GoldG	26.2	24.3	16.6	22.1	27.7	17.8	11.0	14.8	24.0	HF Checkpoints 🤗
YOLO-World-M	O365+GoldG	31.0	28.6	19.7	26.6	31.9	22.3	16.2	19.0	28.7	HF Checkpoints 🤗
YOLO-World-L	O365+GoldG	35.0	32.5	22.3	30.6	36.1	24.8	17.8	22.4	32.5	HF Checkpoints 🤗

NOTE:

1. The evaluation results of AP^fixed are tested on LVIS minival with fixed AP.
2. The evaluation results of AP^mini are tested on LVIS minival.
3. The evaluation results of AP^val are tested on LVIS val 1.0.

Getting started

1. Installation

YOLO-World is developed based on torch==1.11.0 mmyolo==0.6.0 and mmdetection==3.0.0.

python setup.py build develop

2. Preparing Data

We provide the details about the pre-training data in docs/data.

Training & Evaluation

We adopt the default training or evaluation scripts of mmyolo. We provide the configs for pre-training and fine-tuning in configs/pretrain and configs/finetune_coco. Training YOLO-World is easy:

chmod +x tools/dist_train.sh
# sample command for pre-training, use AMP for mixed-precision training
./tools/dist_train.sh configs/pretrain/yolo_world_l_t2i_bn_2e-4_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py 8 --amp

NOTE: YOLO-World is pre-trained on 4 nodes with 8 GPUs per node (32 GPUs in total). For pre-training, the node_rank and nnodes for multi-node training should be specified.

Evaluating YOLO-World is also easy:

chmod +x tools/dist_test.sh
./tools/dist_test.sh path/to/config path/to/weights 8

NOTE: We mainly evaluate the performance on LVIS-minival for pre-training.

Deployment

We provide the details about deployment for downstream applications in docs/deployment. You can directly download the ONNX model through the online demo in Huggingface Spaces 🤗.

Demo

We provide the Gradio demo for local devices:

python demo.py path/to/config path/to/weights

Acknowledgement

We sincerely thank mmyolo, mmdetection, GLIP, and transformers for providing their wonderful code to the community!

Citations

If you find YOLO-World is useful in your research or applications, please consider giving us a star 🌟 and citing it.

@article{cheng2024yolow,
  title={YOLO-World: Real-Time Open-Vocabulary Object Detection},
  author={Cheng, Tianheng and Song, Lin and Ge, Yixiao and Liu, Wenyu and Wang, Xinggang and Shan, Ying},
  journal={arXiv preprint arXiv:2401.17270},
  year={2024}
}

Licence

YOLO-World is under the GPL-v3 Licence and is supported for comercial usage.