search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.38k stars 650 forks source link

MSRA-TD500数据集的config文件能否公布下呢 #537

Closed 6098669 closed 2 years ago

6098669 commented 2 years ago

这个数据集太小了一直很难复现出ABCNetV2中的结果，跪求MSRA-TD500的config文件

Yuliang-Liu commented 2 years ago

@6098669 Did you use the Chinese pretrained model to finetune on MSRA-TD500? The config for MSRA-TD500 as well as the models will be scheduled to release.

ocrhei commented 2 years ago

同求MSRA-TD500数据集和config文件

mxin262 commented 2 years ago

I reproduced the results of the MSRD-TD500. You can reproduce it as follow.

(1) According to the paper, directly use the official Chinese pretrained model. f7e12db1f1417e23244452fd0ee9b3e

(2) Change the config as follow. e722f541a99184296e753b7bac8ded6 And you also need to change the config to use the training data of MSRA-TD500.

(3) Add the path of MSRA-TD500 here. The train&test data can be downloaded from BaiduNetDisk Passwd: b8ki

(4) Because the MSRA-TD500 does not involve recognition training, the following modifications are made here. image image

(5) Train the model. OMP_NUM_THREADS=1 python tools/train_net.py --config-file configs/BAText/TotalText/v2_attn_R_50.yaml --num-gpus 2 MODEL.WEIGHTS model_v2_chn_pretrain.pth

(6) This is the result of my reproduction, which is consistent with the paper. image