search

zhang-tao-whu / DVIS

DVIS: Decoupled Video Instance Segmentation Framework
MIT License
127 stars 6 forks source link

A problem of training R50 model under VIPSeg dataset #9

Closed xihuachen9804 closed 1 year ago

xihuachen9804 commented 1 year ago

Hi! Your work is excellent! I try to train DVIS_R50 model under VIPSeg dataset. Do I need to finetune the segmenter by following "Training on a new dataset"? Or I just need to train by following "Training" and use minvis_pretrained_weights.pth like "minvis_ovis_R50.pth"? I‘m looking forward to your answer. Thanks!

zhang-tao-whu commented 1 year ago

Yes, you need to finetune the segmenter on the vipseg dataset. I will release the weights of the fine-tuned segmenter, please wait a few hours.

zhang-tao-whu commented 1 year ago

weight.pth, the password is 'dvis'.

xihuachen9804 commented 1 year ago

weight.pth, the password is 'dvis'.

Thank you so much for your sharing!

zhang-tao-whu commented 1 year ago

MinVIS_VIPSeg_R50.pth and MinVIS_VIPSeg_SwinL.pth are released. Please refer to here.

yuyuan2 commented 1 year ago

MinVIS_VIPSeg_R50.pth and MinVIS_VIPSeg_SwinL.pth are released. Please refer to here.

Hi, I have a question about testing result on VIPSeg, I inference with VIPSeg Swin-L weights but get a VPQ value of around 52, which is much lower than 57.6. Could you pls tell me how to calculate the reported VPQ value?

I also would like to know how did you calculate VPQthing and VPQstuff.

Thanks

zhang-tao-whu commented 1 year ago

Which model did you use to predict the results? The accuracy achieved by the prediction results of 57.6 VPQ is based on DVIS-Offline (SwinL).

yuyuan2 commented 1 year ago

Hi, actually I used DVIS-Offline (SwinL) model provided in MODEL.ZOO, I would like to know if the 57.6 VPQ are based on test set?

zhang-tao-whu commented 1 year ago

The predicted results for the thing and stuff can be found in the vpq-final.txt file after using the evaluation script eval_vpq_vspw.py provided by VIPSeg.

zhang-tao-whu commented 1 year ago

The accuracy presented in the DVIS paper refers to the validation set, for the accuracy on the test set, please refer to our championship technical report at vps report.

yuyuan2 commented 1 year ago

Thanks much for your information!

yuyuan2 commented 1 year ago

Hi, Zhang,

I have another question,

For VSPW dataset, which is the eval.py file to get mIoU performance? Did you also report on val dataset?

Thanks

zhang-tao-whu commented 1 year ago

I haven't tested the accuracy of the result on VSPW dataset locally. I only submitted the results to the competition server. Maybe you can find the evaluation code in VSPW's GitHub repository.