[VLDB'20] Panorama: A Data System for Unbounded Vocabulary Querying over Video

[ganler]

Project Home Page: https://adalabucsd.github.io/panorama.html

[ganler]

Vocabulary

e.g., GoogleNetCar can recognize cars in over 400 species.

In-vocabulary recognition: There're pre-trained models to recognize a specific kind of object. Out-vocab recognition: No model specifically trained for that.

Challenge

• The user owns a black-box model and he or she may want to extend the model's vocabulary.
• Speedup.

Traditional Approach

Use expertise & re-training for an extension.

Panorama

Use PanoromaNet as a substitute for the black-box model( with extended vocabulary).

Contrib.

Conclusion

Not too many system insights, but AI stuffs.