C4W3 Yolo Algorithm

Hi there,

Recently, I have spent some time to study Yolo v1 and plan to modify it to implement a specific goal: detect car’s license plate.

But I am NOT sure if my idea will work in theory, so I present my thought as below, would you pls help review it and give me any suggestion?

Modification on net structure

The code of Yolo v1 I got is design to detect 20 classes objects. the last two layers are:

dense_2 (Dense) (None, 1470)
yolo_outputs (YoloOutputs) (None, 7, 7, 30)

Due to images I will use to train the model always includes a license plate, I plan to move class information in the output and change the last 2 layers as follows:

dense_2 (Dense) (None, 490)
yolo_outputs (YoloOutputs) (None, 7, 7, 10)

Modification on loss function

the original loss:

loss = class_loss + (0.5 * no_object_loss + object_loss) + 5 * box_loss

Because of the same reason, my version:

loss = (0.5 * no_object_loss + object_loss) + box_loss

is there anyone to study yolo v1?

The original YOLO paper was published in early 2016. The second paper, YOLO 9000, also came out in 2016 and addressed shortcomings in the first version. Since then, there are multiple other releases from the YOLO team as well as advances made by others in the field. Are you sure you don’t need any of those improvements and enhancements?

thanks for your reply.

I am a beginner, just want to use this mini project to do some practice on CNN & tensorflow and try to understand deep learning more deeply, NOT to solve a real problem.

In the all yolo versions, yolo v1 could be easiest one for me, so I choose it to do more study.

1 Like

Fair enough. Simple is always a good place to start. Depending on the training data, you may expose challenges that lead to the later modifications and see for yourself why they were adopted.