How to reduce background image in ReIdentificaiton

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Please provide complete information as applicable to your setup.

**• Hardware Platform (Jetson / GPU)**RTX4080
• DeepStream Version7.1
• JetPack Version (valid for Jetson only)
• TensorRT VersionTensorRT-8.6.1.6
i am using ReIdentification Net from Tao in deepstream-app. PGIE is peoplenet and SGIE is ReIdentification Net.
Found out that sometimes background makes dilution of the feature vector. How can I reduce the effect?

3

ReIdentificationNet was trained using ordinary rectangular person crops from Market-1501 and AI City/MTMC—not foreground-masked images. Replacing the background with black pixels creates a different input distribution and sharp silhouette boundaries that the model did not see during training.

  1. What is your deployment scenario—indoor, outdoor? Could you provide examples showing how the background affects matching? do you want to blur the background or set pixels outside the mask to a fixed value?
4

Any update? Thanks!
A feature vector is a numerical embedding, so could you clarify how you measured or observed the background “diluting” it? Please provide example person crops, similarity scores, or matching failures if possible.
ReIdentification is a pretrained model. You can retrain it with representative data so that embeddings of the same person remain similar across different backgrounds.

5

Yeah thanks I’ll get back to you.