search

WongKinYiu / ScaledYOLOv4

Scaled-YOLOv4: Scaling Cross Stage Partial Network
GNU General Public License v3.0
2.02k stars 574 forks source link

Training/Testing with COCO2017 Dataset #209

Open hojoungjang opened 3 years ago

hojoungjang commented 3 years ago

How do I use this repo with COCO dataset? I don't think you guys have datapipeline. What can I do to prepare COCO dataset to run training or testing?

hojoungjang commented 3 years ago

When I run training with pre-trained weights loaded and raw COCO dataset (no preprocessing, downloaded only), all the bounding box predictions are off by a fixed amount with correct class predictions. So my guess is that the predicted bounding box coordinates are off because images and bounding box coordinates are not adjusted properly. Any idea?

WongKinYiu commented 3 years ago

https://github.com/AlexeyAB/darknet/blob/master/scripts/get_coco2017.sh

annotation of yolo use x_center and y_center while other methods usually use x_top_left and y_top_left, i guess it is the reason.

the00700 commented 8 months ago

i too have an issue, i wrote my own machine learning code to try and train the coco2017 dataset with my own take on a faster CNN model but my predictions are way off, i get values like [ 3.2514e-05, -1.3630e-02, 1.6813e-02, 2.1317e-02] when i am supposed to get values like [241.14,43.38,44.74,91.96] any idea why?