- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi !
I'm pretty new concerning NN and I wanted to know if I'm following a right path for my application. Moreover, I want to learn how the Yolo algorithm and a caffe model work through examples.
My objective is to adapt the tiny yolo example with my own caffemodel to use on the Movidius neural stick. For this, I want to fine-tune (as seen in this official caffe example) a ImageNet ilsvrc pretrained model .
Concerning the application, I need to conceive the computer vision system for a robot that need to collect bottles and avoid some obstacles. I need the following classes : Bottle, grass, rock/gravel and cinderblock. The obstacle are the cinderblock, the grass and rock/gravel are special kind of ground that the robot might enconter and the bottle speak from itself.
My concerns are :
- Can I use the data from ImageNet to fine-tune the network (datas for the bottle class) ? Or is this inefficient since the pretrained model use those datas ? Do you know where I can find more labeled dataset ?
- Will it work? I'm sorry if this is a dumb question. Before going through those two examples in entire, I wanted to know if those are the right example to read to realise the application.
Thanks.
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@gringet You should be able to use any pictures of bottles for training if that is what you want to detect. You can find your own images and label them yourself for your dataset. Although it isn't TinyYolo or object detection, AshwinVijayakumar has a blog about deploying a custom model which involves taking GoogLeNet and retraining it on a different data set @ https://movidius.github.io/blog/deploying-custom-caffe-models/.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
@gringet, in addition to what @Tome_at_Intel already recommended, here are some tips to prep your own dataset:
- For object detection, http://image-net.org/challenges/LSVRC/2015/ui/det.html might be a good resource. However, I couldn't find a search criteria specifically for 'bottle'.
- Alternatively, you can try to use the CLS-LOC dataset (generic ImageNet dataset) to train your detection model.
- Head over to http://image-net.org/challenges/LSVRC/2017/browse-synsets, make note of the Synset words you are interested in, and download the images using the steps described in http://image-net.org/download-imageurls
- In order to get feasible accuracy, you may have to train your model on a 'true object detection' dataset first, and then fine-tune it on the CLS-LOC dataset.
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page