search

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).
MIT License
1.73k stars 784 forks source link

Display target data #395

Open zoecarver opened 5 years ago

zoecarver commented 5 years ago

I am wondering how to display the target data. The BatchGenerator does a lot of preprocessing to the data so when I try to load an element from y_batch the boxes are all messed up. My guess is that this is because of how it decodes the boxes (specifically these lines).

Training still seems to work even though the target appears to be off, why is this? Also what is the best way to counter the decoding?

robertlugg commented 5 years ago

I decided to only check during inferencing time. You might edit predict.py. yolo.predict() is called. From the returned boxes, just multiply by image size to convert from relative position to pixel locations. For the 'true' coordinates, I read them directly from the annotation file.

I too am confused by the encoding. It appears to be incorrect with respect to the original paper. However, I don't understand the implementation well enough to say that very strongly :| Plus the fact that this network works suggests that something is happening that I don't understand.

mavala97 commented 4 years ago

@robertlugg @zoecarver Now struggling with the exact same thing, anyone figured it out yet? @experiencor Can you explain the encoding and decoding? Whenever I put y_batch into the netout decoder my boxes are jacked up. And another question, the network output should be the same as ybatch right(from the batch generator)? And if not, what values should y_batch contain(xmin,ymin,xmax,ymax)?