CentripetalText
CentripetalText copied to clipboard
Published
20 hours ago •
shengtao96
shengtao96
CTW1500 训练结果与论文中的结果相差比较大
trafficstars
作者你好，用这个repo中的代码在CTW1500数据集上直接进行训练，参数配置和config/ct下的tt基本相似，复现出的最佳模型指标结果与论文中的结果相差比较大：
论文中：
P: 85.5, R: 79.2, F:82.2
复现结果： p: 0.8434, r: 0.7795, f: 0.8102
请问可以把total-text ctw1500 msra-td500 SynthText的7个最佳模型文件上传一下吗？这样方便进行实验对比和算法再优化，也可以提升引用
期待您的回复，非常感谢！
ctw1500使用的config文件如下：
{
"model": {
"type": "CT",
"backbone": {
"type": "resnet18",
"pretrained": true
},
"neck": {
"type": "FPEM_v1",
"in_channels": [
64,
128,
256,
512
},
"detection_head": {
"type": "CT_Head",
"in_channels": 512,
"hidden_dim": 128,
"num_classes": 3,
"loss_kernel": {
"type": "DiceLoss",
"loss_weight": 1.0
},
"loss_loc": {
"type": "SmoothL1Loss",
"beta": 0.1,
"loss_weight": 0.05
}
}
},
"data": {
"batch_size": 22,
"train": {
"type": "CT_CTW",
"split": "train",
"is_transform": true,
"img_size": 640,
"short_size": 640,
"kernel_scale": 0.7,
"read_type": "cv2"
},
"test": {
"type": "CT_CTW",
"split": "test",
"short_size": 640,
"read_type": "cv2"
}
},
"train_cfg": {
"lr": 0.001,
"schedule": "polylr",
"epoch": 600,
"optimizer": "Adam"
},
"test_cfg": {
"min_score": 0.88,
"min_area": 16,
"bbox_type": "poly",
"result_path": "outputs/submit_ctw/"
}
}
