pavancm / CONTRIQUE

Official implementation for "Image Quality Assessment using Contrastive Learning"
126 stars 13 forks source link

The difference between trainning from image-classification task pretrained model and training from the start #15

Closed jczhydy closed 1 year ago

jczhydy commented 1 year ago

When training, I find training with pretrained resnet50 model takes less trainning time.I want to know why the article choose to training from the start without pretrained?

pavancm commented 1 year ago

Pretrained resnet50 is a supervised trained model which requires object classification labels as ground truth for training. There is a limitation on dataset size based on how easy it is to obtain accurate ground truth labels for object classes. On the other hand this work focuses on self-supervision where ground truth labels can be generated without manual supervision, thus making it easy to create large datasets. We also showed that these self-supervised models achieve competitive performance. The objective was to show that supervised data is not essential for achieving superior performance.

jczhydy commented 1 year ago

Thank you. And I want to learn more about MS colorspace,I failed to get more information when searching for it on the internet. When I train the model only through MS colorpace by the code provided in this website,I can't get a good result like other colorspace like RGB,Grayscale.