search

sfzhang15 / RefineDet

Single-Shot Refinement Neural Network for Object Detection, CVPR, 2018
Other
1.43k stars 393 forks source link

Training with negative images #167

Closed nerius59 closed 5 years ago

nerius59 commented 5 years ago

Hi, thank you for your great job and repo!

I want to know it is possible to work with negative images(images without a true object = images that have no ground truth in its xml file) for training RefineDet.

In order to reduce false-positive detections, I added negative images to my dataset. But it seems that the negative images didn't affect the training result.

I already understand it doesn't work in SSD-caffe without modifying. (like here https://github.com/weiliu89/caffe/issues/146 , and here https://github.com/weiliu89/caffe/issues/145)

It is same for your Refindet? Do you have any good ideas for this matter? Thanks.

sfzhang15 commented 5 years ago

@nerius59 Yes, Refinedet has the same problem as SSD, since our codes are based on SSD. Training with negative images dose not work without modifying, you can refer to SSD issues to modify some codes to support.

nerius59 commented 5 years ago

Thanks. I'll try to modify the code.