MCG-NJU / MixFormer

[CVPR 2022 Oral & TPAMI 2024] MixFormer: End-to-End Tracking with Iterative Mixed Attention
https://arxiv.org/abs/2203.11082
MIT License
461 stars 75 forks source link

Stage 2 training #101

Open samueleruffino99 opened 11 months ago

samueleruffino99 commented 11 months ago

Hello! I would like to train stage 2 only, where can I find the pertained model for stage 1? Apparently on the drive there are online models only for convmae. Thank you!

yutaocui commented 11 months ago

Hi, there are all models of MixViT/MixViT(ConvMAE) and MixCvT on https://drive.google.com/drive/folders/1C0oECsbTB3uWk2MjrxDuVhIazcT975xb.

samueleruffino99 commented 11 months ago

But under MixViT-ConvMAE I can download online models only (already with SPM integrated).

yutaocui commented 11 months ago

You can just drop the SPM module since this module is trained on a separate stage with other weights getting detached.

samueleruffino99 commented 11 months ago

Ok, but then at stage 2 SPM only has been trained ? No more training on backbone and box head ?

yutaocui commented 11 months ago

Yes.

samueleruffino99 commented 11 months ago

But in this way the SPM would be already pretrained or everything is already implemented in the proper way?

yutaocui commented 11 months ago

We think finetuning the SPM module (excepting for backbone) is enough for filtering reliable samples.