Regarding Action Localization Task

[deepsworld]

Hi @kylemin,

Thank you for sharing your excellent work.

Can you please let me know if you use ground truth tracks and detection boxes for AVA dataset?

[kylemin]

Thanks for your interest in this work!
36.8% is achieved when using the ground-truth person_ids. If we use a simple tracking algorithm (based only on IoUs) for the person_ids, our method produces 31.0%.

Best,
Kyle