shengtao96 / CentripetalText

29 stars 6 forks source link

CTW1500 训练结果与论文中的结果相差比较大 #2

Open AlphaNext opened 2 years ago

AlphaNext commented 2 years ago

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/"
    }
}
shengtao96 commented 2 years ago

麻烦等到周末

AlphaNext commented 2 years ago

好的,期待您的回复,非常感谢!

123cjjjj commented 1 year ago

您好,请问您问题解决了吗?

123cjjjj commented 1 year ago

好的,期待您的回复,非常感谢! 你好,请问关于ctw数据集的处理即dataset/ct/ct_ctw.py这个文件有吗?

AlphaNext commented 1 year ago

您好,请问您问题解决了吗?

没有解决,后面没follow这个论文了,确实没复现出来

123cjjjj commented 1 year ago

你好,关于ctw评估的文件可以公开吗?eval文件夹下面的