Visual Understanding Beyond Naming¶
http://strataconf.com/big-data-conference-ca-2015/public/schedule/detail/40181
Summary¶
If you torture data long enough, it will confess.
“90% of the traffic will be visual data” – Cisco This is “Big Data”, which is too big for humans.
Visual data is “digital dark matter” because it’s noisy, unsegmented, high-entropy, two or three dimentional, so it is difficult to handle.
Visual similarity via semantics¶
Using words to understand/index pictures. Object naming -> Object categorization Image labeling (e.g., with CNNs + ImageNet)
Two problems¶
- Long tails! We have huge amount of data.
- There are many more things in our visual world.
Visual world¶
- Not one-to-one (Visual world -> City)
- The language bottleneck