0

I want to use the kitti dataset and training it with the yolo algorithm. In general, the format of the number of objects and the coordinates of the object in an image for each object are as follows:

 <object-class> <x_center> <y_center> <width> <height>

The question is how can I add different fields to this algorithm and training it?

aysa
  • 42
  • 9

1 Answers1

2

What do you mean under "add different fields to this algorithm"?

If you want to train standart YOLO model, you can't just add such parameters as truncated, occluded etc to label file. Best way I see is to redefine labels from kitti to yolo format.

You can just google for open-source solutions like that.

Maxim Lyuzin
  • 377
  • 3
  • 8
  • The Yolo format has 5 parameters, but in Kitti there are 15 parameters that I need 7 parameters. – aysa May 24 '22 at 04:13
  • which 7 parameters do you need? 1 for class name, 4 for bbox, and 2 for truncation and occlusion? – Maxim Lyuzin May 26 '22 at 10:07
  • I want to estimate the distance and I think need the location – aysa May 27 '22 at 06:07
  • Distance estimating isn't 2D problem like object detection. Of course you can add the specific field to your annotations and change YOLO model to make regression of distance. But the values will be based on your train data only, model will simply overfit to them and will be not able to work in real-world data. If you want to use still 2D approach, take a look at [1](https://openaccess.thecvf.com/ICCV2021_workshops/3DODI), [2](https://arxiv.org/pdf/2103.01468.pdf) and its realization [3](https://github.com/griffbr/ODMD) – Maxim Lyuzin May 27 '22 at 08:42