Open xiaohu2015 opened 6 years ago
I think it's a trade-off. Plus side: different objects overlapping or 'entangled' can be let through. Minus side: false positives for weaker classes on the same object (like in YOLO3 article addendum, in the discussion about COCO AP metric).
I think we shpuld do NMS on the boxes for each class. But using tf.image.non_max_suppression means you treat all classes equally. It may be bad.