search

wanghao9610 / OV-DINO

Official implementation of OV-DINO: Unified Open-Vocabulary Detection with Language-Aware Selective Fusion
https://wanghao9610.github.io/OV-DINO
Apache License 2.0
255 stars 14 forks source link

Multi dataset training and ONNX model #49

Open yuzhengfa opened 1 month ago

yuzhengfa commented 1 month ago

Does the code support training on multiple datasets? Does your paper involve an experiment where O365, GoldG, and CC1M data are trained together? Do you have this training script? Also, may I ask if the model supports converting to onnx? Look forward to your reply.

wanghao9610 commented 1 month ago

@yuzhengfa Hello, I'm sorry for the late response. 1) OV-DINO is based on Detectron2, which supports multi-dataset training. You can change the dataset names in the dataconfig.py. The following is a example:

    dataset=L(get_detection_dataset_dicts)(names=["o365v1_train_ovd_unipro", "coco_2017_train_ovd_unipro"]),

It shows how to gather the O365 and COCO datasets, you can change it by yourself. 2) O365, GoldG, and CC1M are trained together with our Unified Data Integration pipeline. 3) The training script and the logs of [O365, GoldG, and CC1M] will be released after our paper has been accepted, please stay tuned. 4) This repo doesn't support ONNX export yet, I have another project to do and don't have much time to fix it. I will consider adding this feature as soon as possible.