search

lucasjinreal / kitti-ssd

Train your own data using SSD in a more clear and simple way(not include source code)
102 stars 44 forks source link

Quick Question: Must label all interested objects? #19

Closed LeonDong1993 closed 5 years ago

LeonDong1993 commented 5 years ago

When labeling our own dataset, for each image, are we MUST label all the objects in the images that we are interested ?

If object A is in the image but we do not label it, do this means you are telling the Neural Network, the region that object A lies is background. In other words, all the regions in the image that are not labelled is treated as class 'Background' ?

LeonDong1993 commented 5 years ago

@jinfagang Hope you can help me out.

lucasjinreal commented 5 years ago

@LeonDong1993 Definitely..... all classes should be labeled otherwise it suppose to be background

LeonDong1993 commented 5 years ago

@jinfagang Thank you so much, you are so nice !

Just one more question, if I have m ground truth box and the NN give n predicted box. How to evaluate the accuracy in this case ? I know IoU is used to evaluate the precision between two boxes. But how when there are multiple true and predicted box?

LeonDong1993 commented 5 years ago

I have figured it out! Thanks !