search

ZQPei / deep_sort_pytorch

MOT using deepsort and yolov3 with pytorch
MIT License
2.82k stars 726 forks source link

Data set structure mismatch #145

Open farooqzaman1 opened 4 years ago

farooqzaman1 commented 4 years ago

Hi, you used data Loader from pytorch which require all images to be in separate folders each folder name should be a class name, and will contain all images of same class. All datasets you refers to are not in this format. how can i use a new dataset to retrain your model.

Zsugabubus27 commented 4 years ago

Hi!

I also wanted to retrain the model and I have a little bit different dataset where the images are in one folder and the class of each image is encoded in the file name. If you want to create such unique dataset, you have to create a custom dataset class in pytorch. It is very easy if you follow this discussion: How to load Images without using ‘ImageFolder’

WurmD commented 4 years ago

+1 to this issue.

I expect the author transformed the Market1501 dataset to the necessary folder structure with a python script. The addition of this script to the repository would be appreciated.

joaquinCaceres commented 3 years ago

At this link is the MARS dataset in the correct format. The problem is that when training, the accuracy on the validation set in the time 20 does not exceed 1%