avanetten / simrdwn

Rapid satellite imagery object detection
216 stars 153 forks source link

train my own dataset, the format of label #82

Closed ssunlin closed 4 years ago

ssunlin commented 5 years ago

hello,@avanetten I want to train my own dataset, but I don't want to download the COWC dataset because of big size. I wonder what's the format of labels when executing the yolt2 model. my dataset format is like this: images/0.jpg images/0.txt (the label in txtfile is :cls_id, x, y, w, h ) when I execute the yolt2 model, it looks normal, while the accuracy, precision and recall is NAN. So I want to know what's the label format in the txtfile. Screenshot from 2019-09-24 20-31-41

jacquesmats commented 5 years ago

I guess you are running out of images for training. It looks like you have 623 images, and this NaN problem start right after epoch 623.

sunyuxiu commented 3 years ago

Hello! How you change the your datasets to the net work needed?I having been puzzled this questions for a long time and did not find any way to work it out.

hello,@avanetten I want to train my own dataset, but I don't want to download the COWC dataset because of big size. I wonder what's the format of labels when executing the yolt2 model. my dataset format is like this: images/0.jpg images/0.txt (the label in txtfile is :cls_id, x, y, w, h ) when I execute the yolt2 model, it looks normal, while the accuracy, precision and recall is NAN. So I want to know what's the label format in the txtfile. Screenshot from 2019-09-24 20-31-41

edwardnguyen1705 commented 3 years ago

Hi @sunyuxiu Have you found solution?

sunyuxiu commented 3 years ago

@edwardnguyen1705 Yes,it is maybe need the picture and the list name of picture,the bounding boxes of yolo format also needed.Which step do you on?Training or testing?

edwardnguyen1705 commented 3 years ago

@sunyuxiu I already prepared anything. But my image size is 320x320 (actual size 1600x600, but I cropped patches 320x320, and transform the box coordinates to 320x320). I trained yolov5 on this dataset, there is no problem.

sunyuxiu commented 3 years ago

@edwardnguyen1705 yolov5 can support the hug size of image?My image size is hug with 8000x6000 and the object is so small with around 10x30.With a lot of days of searching for the properly net work and find this net can support the hug size.But recently I meet a new error with this net work when testing.https://github.com/avanetten/simrdwn/issues/101.Can you look and help me?

sunyuxiu commented 3 years ago

@edwardnguyen1705 Now I am trying training the image under the format of tensorflow and building the docker.Do you have an good idea about the training the hug size of image as I mentioned above?yolo v5 or this net?

edwardnguyen1705 commented 3 years ago

@sunyuxiu , I have not tried this one so I do not know. But yolov5 is not suitable for your task, it cannot detect multi-scale objects in an image.