Extracting meaning from images.
WHAT THE COMMON TASKS ARE
- Classification: what is this
- Detection: what is here and where
- Segmentation: which pixels belong to what
- Tracking: following something across frames
- Recognition: identifying a specific instance
- Measurement: extracting dimensions or counts
WHAT DETERMINES DIFFICULTY
Variation in lighting, angle, scale and occlusion How similar the classes are How much labelled data exists
WHAT TRANSFER LEARNING PROVIDES
Starting from a model trained on general images, adapting with a smaller set.
WHY THAT MATTERS PRACTICALLY
It makes useful systems achievable with hundreds rather than millions of examples.
WHAT TO COLLECT
Images resembling actual deployment conditions.
WHY THAT MATTERS MOST
Models trained on clean images fail on real ones: poor lighting, motion blur, unusual angles.
WHAT TO TEST WITH
Photographs taken by actual users, on their actual devices.
WHAT TO BE CAREFUL WITH
Labels that are inconsistent between annotators Classes that are rare in the data Backgrounds correlating with classes
WHY THAT LAST POINT
The model learns the background rather than the object.
WHAT TO RUN ON DEVICE WHERE POSSIBLE
Inference, which avoids data transfer and works offline.
