tensorflow-yolo

tensorflow-yolo copied to clipboard

Hi, the prediction demo works great on images with single objects.

However, it seems it cannot detect multiple objects like the original YOLO version. Am I correct??

In theory or rather technically it can. By default it can predict up to 3 boxes per cell. Since there are 7 cells in both vertical and horizontal direction it could predict up to 7 × 7 × 3 objects (147)

I guess you have run the 'demo.py' to generate the prediction. In the function 'process_predicts()', it only select the cell having the highest confidence, i.e., P(object) * IoU * P(class|object), and return the predicted bbox of that cell which means this function can only generate ONE object in the image. However, the model actually can predict multiple objects, you can implement the non-maximal suppression algorithm to generate multiple objects by your own.

@wuhao5688 Yeah exactly, I missed that point. You should also add a threshold for the confidence, so that it will only show those objects where it is confident enough.