A question regarding net-scale-factor and offsets

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

DS 7
dGPU

I’m successfully using a tandem of FaceNet (FaceDetect | NVIDIA NGC) running on PGIE for detection and an ArcFace ONNX model (open_model_zoo/models/public/face-recognition-resnet100-arcface-onnx/README.md at master · openvinotoolkit/open_model_zoo · GitHub) on SGIE for recognition. The latter model produces a 512 float fingerprint vector, which I compare to a database of fingerprints created earlier using Cosine similarity. That worked fine only, after I set a value of 1 for net-scale-factor for the SGIE config and providing nothing special for `offset´.

For the first time in all my attempts using different models I’m having nearly no ambiguities in the recognition results.

Is there any means for knowing, what these values should be ideally for this ONNX model?

[property]
#net-scale-factor=0.0039215697906911373
net-scale-factor=1
#offsets=123.675;116.28;103.53
3

These two parameters are the normalization parameters for the preprocessing. Please refer to the second paragraph in Gst-nvinfer — DeepStream documentation 6.4 documentation

The value of the “net-scale-factor” and “offsets” are decided by the preprocessing algorithm used in the model training. Please consult the engineer who trained the model.

4

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

5

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.