SynthDet v/s Perception package

Hi @ Why are SynthDet scripts so heavily loaded with the code, when we can directly use the UnityEngine.Experimental.Perception.*? Is SynthDet a precursor to Perception package?

I believe this tutorial achieves the same : https://github.com/Unity-Technologies/com.unity.perception/tree/master/com.unity.perception/Documentation~/Tutorial

Thanks!

[mkamalza]

Hi, SynthDet is an older custom project that uses the Perception engine. It was built based on an algorithm and method introduced in an academic paper from Google Cloud AI: https://arxiv.org/pdf/1902.09967.pdf

Thus, SynthDet does not use the new Randomization tools in the latest version of the Perception package, rather, it has a custom placement and randomization algorithm. SynthDet currently produces the best-performing datasets for grocery object detection.

However, using the new Randomization tools and starting from the tutorial you linked, we have found that you can enhance your Randomizations enough (beyond what’s included in the tutorial), with little additional code, to reach similar levels of model performance. We are currently experimenting with this approach.

Moving forward, the new approach based on Randomizers is our recommended way of using the Perception package for more efficient experimentation and extensibility.

[mkamalza]

@mohammedayub44 You can have a look at the SynthDet V2 project which is currently on a branch: https://github.com/Unity-Technologies/SynthDet/tree/SynthDet_V2_RandomizerBased/SynthDet_V2

This project uses the new Randomizers only, and can achieve very similar model training performance to the original SynthDet.