Hello Ai- Re-training SSD-Mobilenet from static images

Hello all

I have already collected a small dataset of static unlabeled images and was wondering the best way to re-train SSD-Mobilenet as it seems most likely to do what I’m after.

The GUI doesnt seem to allow training from a static image.
Should I use detectnet instead?

Why I’m asking questions, I cant seem to figure out how to download a relationships dataset from Open Images, I’m after the ‘wears’ category to try to get a robot arm to play guess who.

I can tell from my initial tests, its going to need to be re-trained on cartoon people :)

Hi @murray, I would recommend using the CVAT tool to label already-captured images. You can export from CVAT in Pascal VOC format. Then create a labels.txt file and use train_ssd.py the same way that you would as if you used my GUI tool.

Thanks @dusty_mv,
Great tutorials by the way.

That’s a really interesting tool.

I was going to use this project for my NVIDIA Ambassador project, so I want to use the Jetson (2g) to achieve every task.

I might try to collect the images again from the live feed.

Opencv looks more light weight to achieve the same task?


Do you mean for DNN object detection? jetson-inference uses TensorRT which will give the best DNN performance on NVIDIA GPUs and Jetson devices.