Human-friendly data collection system for object segmentation

[wkentaro]

The components of this system is

1. Speech recognition
2. Find moving object in pixel-level
3. Track object in pixel-level or roi-level

For 1. @furushchev Do you know good ros node for speech recognition?

For 3. @iKrishneel Do you know good ros node for tracking in pixel-wise? I expect it uses point cloud, because roi-level tracking is common for 2D image: ex ConsensusTracking

---

tl;dr

I and @inabajsk talked about the importance of human-friendly data collection interface for object segmentation. For example:

• Human says to Pepper "Now I teach you this object".
• Pepper says "Ok, what is that object name?"
• Human says "wallet"
• Pepper says "Ok, please show me"
• Human moves object in front of Pepper.
• Pepper do
  • Find moving object
  • Track it
  • Record it

[furushchev]

For speech recognition, I think google speech recognition is a defacto standard. I have simple speech recognition node for it. If you teach me where to commit, I'll send a pull request. Or you can easily try with rwt_speech_recognition. (but it needs chrome)

[iKrishneel]

@wkentaro Do you want to track in 3D or 2D?

[wkentaro]

@furushchev

For speech recognition, I think google speech recognition is a defacto standard. I have simple speech recognition node for it. If you teach me where to commit, I'll send a pull request. Or you can easily try with rwt_speech_recognition. (but it needs chrome)

Could you please send PR to jsk_recognition? (probably jsk_perception or new package)

Is this google speech recognition node? https://github.com/jsk-ros-pkg/jsk_recognition/pull/1249

@iKrishneel

Both is ok for me, but I prefer pixel/point-wise tracking.

[wkentaro]

For 2. @iory Do you know some program to find moving object? I expect you know much about finding object human holding, for imitation learning.

[k-okada]

2017-04-16 18:52 GMT+09:00 Kentaro Wada notifications@github.com:

Could you please send PR to jsk_recognition? (probably jsk_perception or new package)

I would recommend to use https://play.google.com/store/apps/details?id=org.ros.android.android_voice_message, since mic on mobile phone is highly optimized for recognizing human voice.

-- ◉ Kei Okada

[iKrishneel]

@wkentaro Pixel-wise tracking is not common although there are methods that does it is not as competitive as region or part based methods. Besides it also suffers under harsh image conditions. I wrote a tracking algorithm which is a region based tracking where object region is defined by bounding box. I will update this on my github soon.

[wkentaro]

@iKrishneel @k-okada @furushchev Thank you all.

Currently, I selected the use of optical flow and the demo movies are below: HRP2 camera RGB-D: https://drive.google.com/open?id=0B9P1L--7Wd2vaVlHRC10Z2dwQkk Just camera RGB: https://drive.google.com/open?id=0B9P1L--7Wd2vR3ktR0xUSFRMdkk

And the collected images are such like below:

What I found are

• Optical flow works
• Use of depth is not effective because it leads small object image (because of depth range)
• Tracking is not needed because I'd like to collect image when object pose is changed