How to training to custom dataset?

[gustavojoseleite]

I don't understood how to generate the anotation format to training on my dataset, do you have some documentation for help me? Thanks!

[SikandAlex]

Just examine the bounding box format for the BDD100K annotations and for your current dataset.

Masks come in polyline or raw mask format. For the BDD100K dataset you need to represent them as raw PNG masks.

Bounding boxes are defined in multiple formats. You need to find the format your current bounding boxes are in and convert to the BDD100K format. If you had taken the time to take a look inside the BDD100K annotation files you will find:

"box2d": { "x1": 623.59743, "y1": 308.229303, "x2": 635.355495, "y2": 315.088176 }

The bounding boxes are defined in x1, y1, x2, y2 format.

This is a simple data science question that is not specific to this repository. You can use an automated tool like Roboflow to easily convert between annotation formats such as COCO, Darknet etc.