Train centernet with custom dataset

[ashley-mv]

Hi, Could you please add what are the steps required to train this network for a new dataset and classes?

[zzzxxxttt]

1. convert your data and annotation into pascal or coco format.
2. in datasets/pascal.py (or datasets/coco.py):
   • modify VOC_NAMES (or COCO_NAMES) to your class names
   • modify VOC_MEAN, VOC_STD (or COCO_NAMES, COCO_STD) to your dataset mean and std
   • modify self.num_classes to your num of classes
   • modify self.valid_ids to np.arange(1, num_classes+1, dtype=np.int32)
   • modify self.data_dir to your dataset location

If your dataset can not be converted to pascal or coco format, you will need to write your own dataset class.

[ashley-mv]

Thank you.

[GerardWalsh]

How can we use a pretrained model for training on our own dataset?