Can I use the dataset on YOLOv7 and YOLOv4-tiny-CodePudding

Im a complete noob to computer vision.

I have went through the Roboflow tutorial how to train YOLOv7 with my custom dataset using colab, Tutorial: https://www.youtube.com/watch?v=5nsmXLyDaU4 colab notebook: https://colab.research.google.com/drive/1X9A8odmK4k6l26NDviiT6dd6TgR-piOa

In my custom dataset the labels look like this:

1 0.453125 0.6647727265625 0.434375 0.7244318187500001 0.40625 0.7613636359375 0.4109375 0.8153409093749999 0.3828125 0.8835227265625001 0.3828125 0.90056818125 0.409375 0.9176136359375 0.4171875 0.9772727265625001 0.4578125 0.96875 0.5265625 0.90056818125 0.5671875 0.84375 0.578125 0.7642045453125 0.578125 0.6335227265625 0.5625 0.6051136359375 0.54375 0.6022727265625 0.5125 0.6335227265625 0.4671875 0.6448863640625 0.453125 0.6647727265625

this is a lable for a picture with a single class, it has a lot of coordinates because I used Roboflow's editor to draw a precise contour around my image, so it has many "dots" since the contour is made up of many individual lines.

The YOLOv7 net trained with out any problems with ~98% accuracy and was able to id my object very well.

When I tried using the same data set to train YOLOv4-tiny per this tutorial: https://www.youtube.com/watch?v=H3SJcwttTi4 colab notebook: https://colab.research.google.com/drive/1hQO4nOoD6RDxdbz3C1YSiifTsyZjZpYm?usp=sharing

The training failed, it gave only about ~40% accuracy and the training log complained about not having the correct coordinates (I know I should have kept the log to show, my bad)

I realize that the notebooks and the repos are different, but I cant shake the feeling that it might be that YOLOv4-tiny does not like complex geometry labeling, that it wants the box.

I expected that training youlov4-tiny would be as straight forward as training yolov7 with the same dataset. In the end I trained yolov4-tiny with a dataset where the labels were given as a set of 4 coordinates - making a box around the object (unlike the dataset for yolov7 where the labels were given as a set of multiple coordinates because I drew a complex contour around the object)

If anyone had this experience before, I would be grateful to hear your thoughts.

Thank you.

CodePudding user response：

I believe YOLOv4-tiny only supports bounding box annotations at this time. Did you label with polygons?

EDIT: So you're saying you retrained with bounding boxes and you were successful that way?

CodePudding user response：

For more on getting started in computer vision (to beef up Knowledge):

Check out the Jump Start page in your workspace: https://app.roboflow.com/jumpstart. And also: https://roboflow.com/learn