[Recognition] Object classification of new objects (given 45 minutes before the task)

[wkentaro]

Introduction

ARC2017 has new objects revealed 45 minutes before the challenge.
In that case, it is hard to collect training data and train the network, if we use deep learning approach in particular.
Possibly we can do something innovative to solve this problem.

Ideas

Idea1: Class segmentation network

We define object-class using existing objects, which is something more abstract than object name: book, cloth, box, transparent, etc.
And just before the challenge, we can define the object-class for newly passed objects.
By doing this, we can detect/segment objects by their object-class, not object-name.

Idea2: Run multi-gpu

• Measure how many gpus are required @wkentaro

Suggested at jsk-ros-pkg/jsk_recognition#2041 (comment)

[wkentaro]

@yuyu2172 We're planning to release this at 3.2.0, so please try the in-hand recognition with @knorth55 today.

[yuyu2172]

@wkentaro
OK

[wkentaro]

It seems that we cannot make it before 3.2.0, so I postpone this part to 3.3.0.
This is addressed by #2222

[wkentaro]

Resolved by @wkentaro