search

natureyoo / ContinualTTA_ObjectDetection

This is the official code implementation for 'What, How, and When Should Object Detectors Update in Continually Changing Test Domains?' presented at CVPR 2024.
Apache License 2.0
11 stars 0 forks source link

About training sample access #2

Closed sxyxs closed 2 months ago

sxyxs commented 3 months ago

Hi,

Thank you for the great work and congratulations on your paper being accepted by CVPR 2024. When reading your paper, I noticed that you pre-computed a feature distribution from the training set (source domain data) before starting adaptation. Which means, it is not source free. Am I understand it correctly?

natureyoo commented 2 months ago

Yes, obtaining source statistics is easy and requires minimal work during the training phase. Additionally, we only use source feature statistics such as mean and variance to estimate the feature distribution, so the storage required is minimal compared to model checkpoints. Thank you for your interest!

sxyxs commented 2 months ago

Thanks for the clarify!