| Summary: | Move caffe from meta-refkit-extra to meta-refkit | ||
|---|---|---|---|
| Product: | [Yocto Project Subprojects] IoT Reference OS Kit | Reporter: | Ismo Puustinen <ismo.puustinen> |
| Component: | intel-iot-refkit-profile-computer-vision | Assignee: | Ismo Puustinen <ismo.puustinen> |
| Status: | RESOLVED FIXED | QA Contact: | |
| Severity: | enhancement | ||
| Priority: | Medium | CC: | matti-pekka.sivosuo |
| Version: | 2.4 | ||
| Target Milestone: | 2.4 M1 | ||
| Hardware: | x86 | ||
| OS: | x86_64 | ||
| Whiteboard: | |||
| OS type for building Yocto: | --- | Type of Regression: | --- |
| Verified: | Documentation change: | Don't know | |
|
Description
Ismo Puustinen
2017-05-08 12:51:49 UTC
One way forward might be to enable OpenCV DNN module. It can read Caffe models and implements a subset of Caffe layers. http://docs.opencv.org/trunk/d5/de7/tutorial_dnn_googlenet.html http://docs.opencv.org/3.1.0/d6/d0f/group__dnn.html DNN module needs to be tried out in practice first and possibly benchmarked against real Caffe. If it works well enough, it might be enough to solve the Caffe classification issues for most of the users. Real Caffe could then be used only for complex or non-standard models and non-cv use cases, and might continue living in meta-refkit-extra. DNN build fix is in meta-openembedded master-next now. https://patchwork.openembedded.org/patch/140142/ PR for this feature: https://github.com/intel/intel-iot-refkit/pull/186 The DNN module is now enabled and a test is added to Refkit. The test downloads an ImageNet Caffe model (the same used in the meta-refkit-extra Caffe demos) and uses it to classify an image without enabling meta-refkit-extra layer. In the future the caffe recipe could be changed to use the caffe-reference-bvlc recipe for downloading the test data. |