Create tensorflowDetection2D app

[davidvelascogarcia]

Tensorflow: Detector 2D C++ API

Addresses #92.

Description

The pull request adds tensorflowDetection2D module. tensorflowDetection2D module use tensorflow C++ API. The module detects using pre-trained models like VGG16, Inception, ResNet, MobileNet and other pre-trained models. The module build with cmake instead of bazel using tensorflow_cc. Also use YARP to send video source pre and post-procesed. Also admits YARP source video like input. The module has been tested with handstracking model and ssd_mobilenet_v1_coco. This module also publish detection results in YARP port.

C.I

Continuous integration with Travis is in process here. Right now, i´ve problems with Travis timeout. C.I works using Docker image with tensorflow_cc pre-build.

Installation guides

Installation guides done at roboticslab-uc3m/installation-guides#64.

[PeterBowman]

Please add add_subdirectory(tensorflowDetection2D) below this line.

[davidvelascogarcia]

Travis CI: Error

• Travis CI: timeout error
• Travis CI: last test Edit: Bazel Install added.

[davidvelascogarcia]

@PeterBowman New entry added at f01931b.

[davidvelascogarcia]

The last frozen_graph.pb file that i tested was to heavy to be uploaded to github without LFS.
The model was ssd_mobilenet_v1_coco_2018_01_28.
Attached pre-trained models list.

• Tensorflow: Pre-trained models

[jgvictores]

Update to https://github.com/roboticslab-uc3m/installation-guides/blob/master/install-tensorflow_cc.md

[davidvelascogarcia]

Done at 74eb52c.

[jgvictores]

Review to remove OpenCV dependency as well as adquisition/viewer, just in/out ports like

vision/programs/colorRegionDetection2D/ColorRegionDetection2D.hpp

Lines 18 to 19 in 9291fbb

	BufferedPort<ImageOf<PixelRgb> > inImg;
	BufferedPort<ImageOf<PixelRgb> > outImg;

[davidvelascogarcia]

@PeterBowman Thanks. I´m using docker in tensorflow-cpp-examples repo. Right now, I can load docker image with ubuntu and tensorflow_cc. I´m trying to install yarp on it.

[davidvelascogarcia]

@PeterBowman @jgvictores Travis CI working with tensorflow_cc docker image, yarp and opencv in tensorflow-cpp-examples repo test.
Right now, build takes 15 minutes but without using cache because i don´t know how to make a cache of docker image data.

[PeterBowman]

If I'm correct, a new Docker image can be started so that everything (tensorflow_cc + YARP + OpenCV etc.) will be preinstalled, we'd just need to git pull the vision repository. @jgvictores I'd create a roboticslab organization at Docker Hub in which our own images can be hosted. Let's speak 💀2💀?

[PeterBowman]

I'd create a roboticslab organization at Docker Hub in which our own images can be hosted. Let's speak skull2skull?

Outcome:

• https://hub.docker.com/u/roboticslabuc3m
• https://hub.docker.com/u/asrobuc3m

[davidvelascogarcia]

SegmentorThread added at f3881d8, fix format and similar to haarDetection2D.

[davidvelascogarcia]

tensorflowDetection2D trained model and label map model drive link added at 55256a4.
I have added .md at /share/tensorflowDetection2D/conf path instead of .pb and .pbtxt with small installation guide because i didn´t know where to put the explanation about how to set the models.

[davidvelascogarcia]

Readme moved to path program at 54891dd

[davidvelascogarcia]

Waiting for revision.

[jgvictores]

1. I'm seeing that the README lacks some info, such as the algorithm itself (looks like a YOLO or similar). Some info from this PR description could be used/ported.
2. labels_map.pbtxt is in the PR but also in the Google Drive link. I'd remove it from the PR but we should also be careful and remove it from CMakeLists.txt.

[davidvelascogarcia]

• labels_map.pbtxt removed at e148a95 and 17fa21b
• Some info about tensorflowDetection2D at vision-fork