search

Alibaba-MIIL / ImageNet21K

Official Pytorch Implementation of: "ImageNet-21K Pretraining for the Masses"(NeurIPS, 2021) paper
MIT License
724 stars 71 forks source link

could I transferred the paper's idea to object detection task? #44

Closed cs-heibao closed 3 years ago

cs-heibao commented 3 years ago

Recently, do an object fine-grained detection task, is it convenient to transferred to object detection? thanks

mrT23 commented 3 years ago

in the revised version of the paper, i share results on ms-coco detection task, with MTResNet backbone, where switching pretraining from 1K to 21K improved results.

some people said in the git that with resnet50 they were not able to get an improvement. i dont know, i suspect they might had a bug in loading the weights., since resnet50 is a crappy network anyway, my recommendation is not to use it, and switch to mtresnet (or MBV3). the original 1K weights can be found at: https://github.com/Alibaba-MIIL/TResNet/blob/master/MODEL_ZOO.md