What is the performance after training using this repo code?

[ghimiredhikura]

Hello @AI-liu,

You have done wonderful work. The pretrained model you provided is based on DarkNet training. How about training with this repo pytorch code. Is is incomplete? I was training with your code, but couldn't get good results. Please share your results.

Thanks.

[Hizhaoyuan]

Hi, you can take a look at this project, which has a complete implementation and effect presentation.https://github.com/wwooo/tensorflow_complex_yolo

[ghimiredhikura]

hi @Hizhaoyuan,

Thanks for the link and it seems you have done great job.
I have some questions though. While training it looks like you haven't used angle information in anchor box. As well as while comparing iou between two box i think you did not compare rotated box. Isn't is necessary to include angle while training.

I am also working on this and I am getting similar result as yours. The performance is better in hard examples compared to easy and moderate. I was little confused as I was expecting largest AP in easy samples compared to hard samples. Any thought on this?

Best,
Deepak

[Hizhaoyuan]

Hi @ghimiredhikura

When matching anchor and ground truth, the main role of iou is to find an anchor with a close center and similar shape for each ground truth, there is no need to consider the yaw Angle of ground truth. Neural networks can return information about angles. In this case, the neural network simply returns two real numbers, the real part and the imaginary part of the complex number to determine the Angle, which is easy for the neural network. In addition, if iou is calculated with rotated boundingbox, when the shapes of the two bounding boxes are exactly same, but the directions are opposite, the calculated iou is still equal to 1.

The logic of kitti evaluation script is complex. It will automatically filter out some test results that don't care about. This is just an approximate result., you can study the official evaluation script carefully.

[ghimiredhikura]

Thanks @Hizhaoyuan,