Small NNs to reuse the pytorch impl to do my own Object detector?

1

I have read that YOLO, Retina with Resnet, SSD, etc can do Object Detection.

Now I struggle Implementing these from scratch, so I wonder if it is possible to re train some implementation of these from Pytorch, but with a new dataset not included in the pre-trained classes.

I am looking for some NNs that would run in not so powerful devices, so light ones.

Im kinda directionless atm.

2

Yes, fine tuning a model with a new dataset is possible.
torchvision provides implementations for RetinaNet and SSD.
The Object detection finetuning tutorial could be a good starter.

3

But I am trying to find a small model like tiny yolo or any ssd mobile net that does object detection, is this feasible as well? I dont really see anything similar in the docs, after days reading the all other websites as well. @ptrblck

4

Maybe SSDLite would work for you. I don’t think Yolo is available in torchvision.

1 Like
5

Do you now the reason or if it will be ? Seems like an interestin architecture and quite successful

6

I believe it was planned and discussed here but it seems the author is busy with other work:

The only bad news for us is that I transferd to a new work group and my new leader only allow me to share only a half day per week to complete this job. So it might takes me more time than expected.