Accurate in validation but got wrong result on new test data

[Barry-Liang]

I have generated a custom dataset using the object data tools mentioned in this repository. After validation, the final result is as follow:

I think this result is very good. However, I want to use it in a real-time detection case. When I use the same camera and object to detect new data, the test results are totally wrong.

I am sure that the program I used to test the new data is correct. Because When I use it to test the original dataset, the results are all correct. But for the new data, the results are all wrong.

Do you have any idea about what happened?
I guess since the video is captured continuously, all the images in the test dataset have similar images in the train set, therefore the final result is still overfitted. But I am not sure.
I will appreciate it if you can help me with that.

[ValiaVl]

Hi @Barry-Liang . Can you share the code for visualizing the results on new unseen images without the ground truth labels required?