Centernet: how to train model on a few classes

Created on 8 May 2019 · 1Comment · Source: xingyizhou/CenterNet

I don't need all 80 classes from COCO. I need only 2 of them. And my own dataset consists of images of 2 classes. How to train ctdet in such a case if I really want to remove other classes from training process?

Source

Arfua

👍1

Most helpful comment

Hi,
Basically there are three steps to add new dataset:

Convert the dataset annotation to COCO format. Please refer to src/tools/convert_kitti_to_coco.py for an example to convert kitti format to coco format.
Create a dataset intilization file in src/lib/datasets/dataset. In most cases you can just copy src/lib/datasets/dataset/coco.py to your dataset name and change the category information, and annotation path.
Import your dataset at src/lib/datasets/dataset_factory.

xingyizhou on 8 May 2019

👍5

>All comments

Hi,
Basically there are three steps to add new dataset:

Convert the dataset annotation to COCO format. Please refer to src/tools/convert_kitti_to_coco.py for an example to convert kitti format to coco format.
Create a dataset intilization file in src/lib/datasets/dataset. In most cases you can just copy src/lib/datasets/dataset/coco.py to your dataset name and change the category information, and annotation path.
Import your dataset at src/lib/datasets/dataset_factory.