0

How quantization for object detection models varies from that of classification models? Since detection models need to handle the bbox coordinates(multiple objects in an input),there must be some scaling trick in quantization.

Knilakshan20
  • 59
  • 6

1 Answers1

0

You can look at SSD of Tensorflow model in the model zoo API here

SSD is single shot detection model, it takes the features of detection when looking overthe image to provide the classification score according to label. This type of application is very useful for multiple types of object detection.

dtlam26
  • 1,410
  • 11
  • 19
  • I think it's FP implementation.Does it support quantized models? or do they published any detection based quantized model – Knilakshan20 Aug 27 '20 at 07:54
  • yes, on the link I gave you, they have published the ssd model and also the code to generate them, but they will be stable for now only. They are trying to implement them to tf2.0 then all of this in tf1.x soon be removed – dtlam26 Aug 28 '20 at 05:37
  • I did the post training quantization using tflite.When doing inference,the class and class score remain same – Knilakshan20 Sep 03 '20 at 08:50
  • post training is not sufficient, you should use quantize aware training by setting quantize delay in the training graph. Have you checked the link I gave you? – dtlam26 Sep 04 '20 at 09:03
  • 1
    Yeah, I observed the post-training quantization results are poor quantize aware training.Thanks for the help – Knilakshan20 Sep 08 '20 at 03:18
  • no problem mate. The hell of quantization is still waiting for you – dtlam26 Sep 08 '20 at 12:21