I was using the tensorflow android example which has a placeholder for the location of a ROI (see tensorflow/examples/android/src/.../Classifier.java), but it is not used. I would like to be able to locate the ROI of the classified object. I looked through the code, but it wasn't obvious how the location could be determined. Does tensorflow support this?
Unfortunately that placeholder is not populated, and actually determining the region of interest of an object is still a hard research problem. We don't currently have a released example of a model that can do that in TensorFlow, but if you want to look through the research, the term to look for is 'localization'.
There's a related question you might find helpful here:
Does Convolutional Neural Network possess localization abilities on images?
来源：https://stackoverflow.com/questions/34406792/how-to-to-find-location-roi-of-a-recognized-object-in-tensorflow