Forums - Decrease in detection scores / accuracy

2 posts / 0 new
or Register
Last post
Decrease in detection scores / accuracy
amith.k
Join Date: 21 Jul 21
Posts: 2
Posted: Sun, 2021-10-17 23:28
Top

Hi all,

I am working with a custom trained (re-trained) ssd mobilenet v2 model from tensorflow model zoo. The re-trained model has an average precision and recall of around 0.5 when inference tested on my development environment. But post conversion to a DLC model and running as an android application, it is returning absurdly low detection scores of below 0.1 as shown below.

 

Model - ssd_mobilenet_v2_320x320_coco17_tpu-8 

Tensorflow Version - 2.5 SNPE version - 1.54.2

Layers:

Input - input_tensor:0,

Output - StatefulPartitionedCall/Postprocessor/BatchMultiClassNonMaxSuppression_classes, StatefulPartitionedCall/Postprocessor/BatchMultiClassNonMaxSuppression_scores, StatefulPartitionedCall/Postprocessor/BatchMultiClassNonMaxSuppression_boxes[/quote]

I am assuming the problem is with model conversion. Any pointers to solving this is much appreciated.

Thanks, Amith

 

  • Up0
  • Down0
Re: Decrease in detection scores / accuracy
weihuan
Join Date: 12 Apr 20
Posts: 270
Posted: Sun, 2022-01-09 17:09
Top

Dear customer,

Thanks for your work on our engine development.

Could you please share the model to us for more deeply analysis?

BTW, you can compare the layer accuracy while execute the SNPE with --debug into snpe-net-run option.

Meanwhile, we're developing the accuracy tool internal and will release outside further.

BR.

Wei

  • Up0
  • Down0
or Register

Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Qualcomm Incorporated or its subsidiaries (“Qualcomm”). The content is provided for informational purposes only and is not meant to be an endorsement or representation by Qualcomm or any other party. This site may also provide links or references to non-Qualcomm sites and resources. Qualcomm makes no representations, warranties, or other commitments whatsoever about any non-Qualcomm sites or third-party resources that may be referenced, accessible from, or linked to this site.