search

google / automl

Google Brain AutoML
Apache License 2.0
6.21k stars 1.45k forks source link

Support of Incremental Object Detection #615

Open rose-jinyang opened 4 years ago

rose-jinyang commented 4 years ago

Hello How are you? Thanks for contributing this project. I have a question. Does this project support class-incremental training for saving training time? Let's suppose I've trained a model on the dataset with K classes for 5 days. If a new class is added into this dataset, should we train a model with the expanded dataset (K+1 classes) from begin? If so, it is so expensive, especially in case of object detection in retail store. That's because a new class of good is added frequently in retail store. Can we train a new model in short time with the original weight on the expanded dataset? I think that this is very important function. Please let me know if you have a willing to implement this. Thanks

fsx950223 commented 4 years ago

Interesting.

jcburnel commented 4 years ago

Do you have any recommended paper on this subject ? The naive way would be to change last layer while freezing all the weights but their might be better approach. This is indeed a very interesting problem, however I don't think it is the main purpose of this project but if you work on this don't hesitate to submit a PR.

rose-jinyang commented 4 years ago

Hello Thanks for your reply. Most object detectors do not handle it despite this function is very important. Of course, I think that's because this problem is much difficult. I don't think there is a development team as powerful as google. I believe that EfficientDet by google can handle this problem. If this function would be implemented, EfficientDet by google will be more useful and powerful than others. I send some recommended papers for this project. https://arxiv.org/pdf/1708.06977.pdf https://arxiv.org/pdf/2003.04668.pdf https://arxiv.org/pdf/2003.06957.pdf https://arxiv.org/pdf/2003.07304.pdf To my knowledge, it looks the second paper the best. Thank you and Regards

rose-jinyang commented 4 years ago

Hi Please let me know if you have a willing to implement this. Thanks