Have you ever wanted a fast/accurate object detection model, but didn't have time/resources to train it?
This repo aims to provide a variety trained models on different objects, which you can use directly with YOLOX framework.
We also train models on different views of the datasets that have bigger/lower object area. For example, if you need a face detector that will be used on mobile front cameras, there may be no need to detect faces that have very small area (very far objects). We can take advantage of this and use a lower input resolution, so inference is very fast.
Models are mainly trained with Open Images Dataset V6, which has 600 classes. To re-create the dataset used in the trained models, refer to dataset.ipynb.
| Model | Activation | Input Resolution | mAPtest 0.5:0.95 |
Area | Weights | Experiment |
|---|---|---|---|---|---|---|
| nano | silu | 128x128 | 74.20 | > 10% | github | github |
| nano | leaky relu | 128x128 | 73.68 | > 10% | github | github |
| nano | silu | 160x160 | 72.72 | > 5% | github | github |
| nano | leaky relu | 160x160 | 71.90 | > 5% | github | github |
| nano | silu | 192x192 | 66.97 | > 1% | github | github |
| nano | leaky relu | 192x192 | 66.21 | > 1% | github | github |
Note: You can try using any model with slightly higher/lower input resolution and will also work fine.
- Make script to reproduce datasets
@article{yolox2021,
title={YOLOX: Exceeding YOLO Series in 2021},
author={Ge, Zheng and Liu, Songtao and Wang, Feng and Li, Zeming and Sun, Jian},
journal={arXiv preprint arXiv:2107.08430},
year={2021}
}