Bounding Boxes are not aligned and appear not in the picture

[mojoee]

Does anyone have experienced problems with the bounding boxes in the visual presentation? I have basically 2 kind of problems after training as shown in the picture: 1)Bounding Box (GT and PR) are not aligned with the real object and 2) bounding boxes appear outside of the image. Can anyone help me ?

[super-zyw]

Do you have enough training data?

[capslocknanda]

can you help me with how to visualize the prediction on an image ?

[mojoee]

the code for the visualization is in the jupyter notebook file called valid.ipynb. You just need to run this file and you will get the visualizations.

[liupenglei]

I have the same problems, do you have solve it? thank you

[nithinvenny07]

I am also facing same problem did anyone found solution??

[Allwinraj]

I am also facing the same issue. for my own data. The Groudtruth also shows the wrong location while in training and testing the model. but the loss is at a minimum range. Anyone had a solution??

[Barry-Liang]

I have the same problem for my own data. My case is a single object detection with classNum=1. Both training and validation work well. But after that, I use the camera to do the real time detection, even when there is no object in the image, the bounding box will appear. In addition, the bounding box can become very large or in wrong position.

[mojoee]

I remember I had to change the intrinsic camera parameters! Does that work?

[Barry-Liang]

Yes, I used the intern realsence camera to collect data and train the network. The validation works well with reasonable accuracy. After that, I use another RGB camera to do the realtime detection. I have changed the intrinsic parameters in the corresponding

[mojoee]

I can't recall how exactly I solved it. And it's been too long. Hope someone else can help

[super-zyw]

I am doing all good till now, what is the exact issue here?

[nithinvenny07]

@mojoee may i know how u have created the labels for the dataset thanks in advance