Some questions about training my own dataset
PiBigStar5712 opened this issue · comments
I am learning the multi-objective tracking project of YOLOv5+Deepsort, but now I have encountered some problems.
The target I need to track is not people or cars but other objects. Now I have trained my own YOLOv5 model according to the tutorial.
When I directly use this model to track through deepsort, I found that the ReID of the same target to be tracked would change when the target was partially occluded. I think this is the reason why my ReID didn't retrain, so I asked the original author of the project.Here is the issue
Can I directly use your project to train ReID and put the weight file with the suffix. t7 into the deepsort project?
I have tried to directly use the train.py file in the deepsort project to train ReID, but since the first epoch, the accuracy rate is 100%. I don't know what's wrong.
If the object I need to track has only one category and only one target in that category, then I should have only one subfolder under the train file and test folder respectively, right?
Thank you for your time！
Sorry for the late response.
Your training set only have one identity of one class?
It would be great to conduct the style transfer or using the data augmentation to simulate the occlusion case.
You may also consider to collect the online dataset as I do in https://www.zdzheng.xyz/files/TMM20.pdf
Sorry for the late response. Your training set only have one identity of one class? It would be great to conduct the style transfer or using the data augmentation to simulate the occlusion case.
You may also consider to collect the online dataset as I do in https://www.zdzheng.xyz/files/TMM20.pdf
Thank you for your reply. I will try to add other targets of the same class to my dataset.I hope the results will change.