Retraining SSD-Mobilenet V2

Robotics & Edge Computing Jetson Systems Jetson TX2
, ,
1

Hello,

I’ve had success retraining SSD-Mobilenet V1 with the help of tutorial from Retraining tutorial

When I tested Mobilenet V1 and V2, I liked the performance of V2 more. In the tutorial we use the command wget to download the base model .pth file for SSD-Mobilenet V1. Is there a similar link for V2 object detection base model? I’d like to retrain it on my own set of classes and import it to tensorflow so I have control over things like drawing bounding boxes. My goal is to use the object detection model for object tracking.

$ wget https://nvidia.box.com/shared/static/djf5w54rjvpqocsiztzaandq1m3avr7c.pth -O models/mobilenet-v1-ssd-mp-0_675.pth

Thank you.

2

Hi @cap, you can check the upstream pytorch-ssd project to see if one is available: https://github.com/qfgaohao/pytorch-ssd

I have only tested it with SSD-Mobilenet-v1 and have found the performance to be good. The Mobilenet-v2 optimizations tend to have a bigger impact on smaller mobile platforms (e.g. mobile phones) where separable convolutions are more necessary.

3

Hey @dusty_nv , I’ll give it another go for trying V1. Before I was getting lots of false positives but that could’ve been due to a number of reasons. I’ll try retraining and see what results I can come up with. I also see the repository has a link to the .pth file for what I’m looking for, I’ll keep that in mind as well.

Thank you!