Overall plan

[marc-hanheide]

Abilities:

• Navigation (@mtlazaro)
  • using RGBD to create laser scanner
• Dialogue (@francescodelduchetto with remote input from @andreavanzo )
• People Perception (@sergimolina, with remote input from @robpinillos )
  • detection and tracking
  • identification / maybe demographics estimation
  • sound localisation
• Object Perception (@MFernandezCarmona )
• UI (@marc-hanheide)
• System integration and deployment (@marc-hanheide)

Tasks

See our old document at https://docs.google.com/document/d/1hbKG8kDjDA8ulWavLRsRMJMjw0T4TgYOE_0dg1rUWzw/edit

Stage 0

• Inspection (@mtlazaro)
• Poster (@marc-hanheide)

Stage 1

• Cocktail Parte (@mtlazaro)
• Speech and Person Recognition (@sergimolina )
• Help-me-Carry (@MFernandezCarmona )
• GPSR (@francescodelduchetto )

Stage 2

• Open Challenge (@marc-hanheide)
  • think about AMI (based on work by @MFernandezCarmona )
• Restaurant (@mtlazaro )
• Advance GPSR (@francescodelduchetto )
• Tour Guide (@sergimolina )

[MFernandezCarmona]

Hi,
About object detection. I managed to make Yolo use python and connect to nao and even recognize some objects/persons. Out of the box it works , more or less.
I had to do some tricks, so I forked the original one to this:
https://github.com/MFernandezCarmona/darknet

Next steps would be retraining the network for our own classes, but that is way out of the capabilities of my graphics card. Suggestions?

[marc-hanheide]

Two options:

• Ask @maxhuettenrauch or someone else of @GeriNeumann s team if we could run training on of the GPU beasts we have (maybe @GeriNeumann has a better idea than darknet for out of the box object recognition anyway).
• Put the training job on either GPU enabled AWS EC2 or GCP computing. I'm happy to spend some money on that. I have created an AWS Account for L-CAS some time ago. It would cost around £1 per hour on a 2 Titan/8 CPU machine if I recall correctly.

[soumenms2015]

does it work on windows platform?

[MFernandezCarmona]

@soumenms2015 I'm afraid not. Please check Readme.md