Can I include images without target objects (a.k.a. no bounding box) into the training dataset?

[yqc924]

Hi, I am working on faster-RCNN to detect one type of object. I created my own training dataset, with half of the images containing the target object, and half of them not.

I am wondering if I could include images without target objects (a.k.a. no bounding box) into the training dataset? Can the well-trained faster-rcnn model predict "no bounding box" for images without target object, or it has to yield at least one box for each image?

Similarly, I am wondering if there is a way to combine image classification task with object detection: i.e. first, check image classified as yes or no? -> if yes, further detect the location/bounding box of the object.

Thanks in advance!

[Mahmoud-Khaled-Nasr]

Hi, I have the same problem! any solution yet?

[ojasvisancheti]

Same problem. Please inform us with some solution

[Zhylkaaa]

I guess FRCNN now supports empty inputs: https://github.com/tensorpack/tensorpack/issues/1256