I utilized this platform (Roboflow) to annotate my dataset and now I’m faced with a challenge. The automation of Roboflow started my labeling number sequence at 0.
This became challenging when utilizing the dataset for initialization on coco128 with 80 classes starting from 0-79.
To integrate my effort deployed from Roboflow, I’m now searching for a solution that would convert what was already labeled to commence at 80.
Why? I am adding what was done in Roboflow to the original dataset’s classes so that it can be efficiently trained via yolov5 and for inference after such processes.
ISSUES CAUSED:
The classes are not being seen in the training initialization process and therefore limits the efficacy of the model dev stages. “Please see image”
You would need to update your annotation files to ensure the index values for all annotations match the classes.
Or, alternatively, merge the coco128 dataset with your custom dataset on Roboflow, and export that way. This will solve the issue of the indexes not showing up when validating your model. However, Roboflow makes the index names alphabetical, so they won’t be in the exact same order as coco128, but they will at least show the results for all classes in the dataset.
I found the solution by merging the datasets which was the best for me, but I encountered and error that was affecting the annotations. This altered some of their names converting this to numbers, I’m a bit apprehensive to continue the merging of the other classes until I can find a solution. I have a small but significant sample size suitable for testing but would prove tedious to redo the annotations.
Thank you for your response @Mohamed and acknowledging my digital presence! indeed your efforts were helpful.
