Hello dear community,
I have been working with Tensorflow, Caffe and Darknet in order to create a custom object detector which can run in the Movidius NCS.
So far, I have been unable to do it following different tutorials, export tools, etc.
Do you have any complete guide to achieve this?
My requirement is a very high precision on the single class detector. For example, with a tensorflow Faster R-CNN network, I was able to detect an object above 20 meters with a 96% of confidence, which is a very good scenario
I would appreciate your guidance here
Thanks in advance
Dear colleagues I've searched a lot and found some pieces of info e.g. for cats and dogs image classification, but I've never seen a step by step for a custom object detection from training to fine tuning, as requested by the colleague above.
@Tome_at_Intel, don't let us down!
@AshwinVijayakumar Thanks for the attention!
Before my request I've found one of your articles: movidius.github.io/blog/deploying-custom-caffe-models/
It was very useful, but I understand it is a image classification which differentiates dogs and cats. Unfortunately this was not enough to allow me taking off with the object detection, considering my modest knowledge about the subject.
So thanks for your guidance so far, I'm looking forward for your new article.
Does anybody could recommend a step by step for the custom object detector?
I've found this link https://github.com/listenlink/caffe/blob/ssd/README.md but it only Works with pre-trainned model, I imagine some steps are missed:
1st How to create a dataset of a custom object as well tag this dataset
2nd How to train it generating the model and then the Graph
3rd How to translate this Graph to the Movidius format
Maybe this sounds a quite simple for the majority here, but it would help a lot the beginners (as me). I believe sharing knowledge is the main reason for this forum existance, right?
Thanks in advance!