search

hyseob / PyMDNet

MDNet PyTorch implementation
Other
420 stars 150 forks source link

ImageNet-VID dataset #37

Open Actmaiji opened 5 years ago

Actmaiji commented 5 years ago

hi,what do you mean the "ImageNet-VID dataset" ? Because I can find the 2015 :http://bvisionweb1.cs.unc.edu/ilsvrc2015/download-videos-3j16.php#vid and the 2017. Can any guy give me a link of the ImageNet-VID dataset ?thanks a lot.

hyseob commented 4 years ago

Hi, @Actmaiji It is the 2015 version. I will add the link that you mentioned to the readme. Thank you :)

Actmaiji commented 4 years ago

thanks a lot.But I still have two quetions:first,when you training on vot ot ImageNet-VID,do you have any method to verify the training results(two classification model) or just verify on online tracking? during training ,the precision reach 0.994(Imagenet),but the online tracking results is worse than 0.971.thanks a lot

hyseob commented 4 years ago

@Actmaiji I verified the pretraining policy using the last 20% of frames for each video as a separated validation set, and added them to the training set for the final release. The online finetuning is just verified on tracking. Thank you :)

Actmaiji commented 4 years ago

Thanks a lots for your replying.Sorry,I coundn't understand what the "trainging set" and the "final release" exactly mean.Do you mean that when n_cycle=50. In # Main trainig loop when i = 49,we added the validation set to the training set, and the model parameters still update? And the final release is the final model?Thank you :)

hyseob commented 4 years ago

The "training set" means the entire data used for pretraining. To fine appropriate hyperparameters (learning rate, number of cycles, etc.) for pretraining, I used 80% of Imagenet-VID as a training set and the rest 20% as a validation set. After finding the hyperparameters, I re-train the model from the scratch using the entire Imagenet-VID as a training set without a separated validation set, which is what the "final release" of the code (pretrain/train_mdnet.py) does.

Actmaiji commented 4 years ago

Thank you. I got it.Did you use the same training strategy on vot training dataset? I used to verify it on tracking the data set, just initializing the first frame and then testing the two classification model.Do you think that will work?

sieumap43 commented 1 year ago

hi,what do you mean the "ImageNet-VID dataset" ? Because I can find the 2015 :http://bvisionweb1.cs.unc.edu/ilsvrc2015/download-videos-3j16.php#vid and the 2017. Can any guy give me a link of the ImageNet-VID dataset ?thanks a lot.

The link just died recently. Does someone have the dataset or an alternative link?