DBnet 使用ploy方式，检测弯曲文本，有没有预训练模型呀

[Tomhouxin]

由于样本数据比较少，只有400张，单纯的用resnet18模型做预训练模型太难收敛了，如果用quad没有问题，但是样本大部分是弯曲样本，所以来请教一下，谢谢！

[mm-assistant[bot]]

We recommend using English or English & Chinese for issues so that we could have broader discussion.

[gaotongxiao]

We have added it to our backlog and will release it when it's ready. But for now, we don't have the checkpoint for this setting.

[Tomhouxin]

Ok, Thanks!

How much data is required to use the ploy mode？

[gaotongxiao]

Generally, text detection models can achieve satisfactory performance with not quite a large amount of data. For example, ic15 only contains 1000 training samples.

[Tomhouxin]

这些是训练日志，目测貌似不再收敛了，hmean-iou:hmean停留在了0.8，loss也不在下降了，但是检测结果来看是乱起八糟的。是否要调整学习率继续训练呢？

{"mode": "val", "epoch": 69, "iter": 80, "lr": 0.00482, "hmean-iou:recall": 0.76316, "hmean-iou:precision": 0.84466, "hmean-iou:hmean": 0.80184} {"mode": "train", "epoch": 70, "iter": 5, "lr": 0.00478, "memory": 6697, "data_time": 1.78142, "loss_prob": 0.8945, "loss_db": 0.14975, "loss_thr": 0.41529, "loss": 1.45953, "time": 2.87308} {"mode": "train", "epoch": 70, "iter": 10, "lr": 0.00478, "memory": 6697, "data_time": 0.36489, "loss_prob": 0.73329, "loss_db": 0.13053, "loss_thr": 0.38303, "loss": 1.24684, "time": 1.38242} {"mode": "train", "epoch": 70, "iter": 15, "lr": 0.00478, "memory": 6697, "data_time": 0.27795, "loss_prob": 0.79045, "loss_db": 0.14272, "loss_thr": 0.39014, "loss": 1.32331, "time": 1.27847} {"mode": "train", "epoch": 70, "iter": 20, "lr": 0.00478, "memory": 6697, "data_time": 0.21064, "loss_prob": 0.83678, "loss_db": 0.14476, "loss_thr": 0.42282, "loss": 1.40436, "time": 1.1707} {"mode": "val", "epoch": 70, "iter": 80, "lr": 0.00478, "hmean-iou:recall": 0.7807, "hmean-iou:precision": 0.84762, "hmean-iou:hmean": 0.81279} {"mode": "train", "epoch": 71, "iter": 5, "lr": 0.00475, "memory": 6697, "data_time": 1.92355, "loss_prob": 0.81302, "loss_db": 0.13625, "loss_thr": 0.42131, "loss": 1.37059, "time": 3.07605} {"mode": "train", "epoch": 71, "iter": 10, "lr": 0.00475, "memory": 6697, "data_time": 0.35925, "loss_prob": 0.90666, "loss_db": 0.15217, "loss_thr": 0.44795, "loss": 1.50678, "time": 1.28993} {"mode": "train", "epoch": 71, "iter": 15, "lr": 0.00475, "memory": 6697, "data_time": 0.26736, "loss_prob": 0.94919, "loss_db": 0.16886, "loss_thr": 0.41646, "loss": 1.53451, "time": 1.21419} {"mode": "train", "epoch": 71, "iter": 20, "lr": 0.00475, "memory": 6697, "data_time": 0.2183, "loss_prob": 0.75486, "loss_db": 0.13175, "loss_thr": 0.4056, "loss": 1.2922, "time": 1.27543} {"mode": "val", "epoch": 71, "iter": 80, "lr": 0.00475, "hmean-iou:recall": 0.81579, "hmean-iou:precision": 0.83784, "hmean-iou:hmean": 0.82667} {"mode": "train", "epoch": 72, "iter": 5, "lr": 0.00472, "memory": 6697, "data_time": 1.85821, "loss_prob": 0.82848, "loss_db": 0.14779, "loss_thr": 0.39736, "loss": 1.37362, "time": 3.30308} {"mode": "train", "epoch": 72, "iter": 10, "lr": 0.00472, "memory": 6697, "data_time": 0.54173, "loss_prob": 0.95, "loss_db": 0.16089, "loss_thr": 0.43677, "loss": 1.54766, "time": 1.33972} {"mode": "train", "epoch": 72, "iter": 15, "lr": 0.00472, "memory": 6697, "data_time": 0.22638, "loss_prob": 0.75431, "loss_db": 0.13126, "loss_thr": 0.3615, "loss": 1.24707, "time": 1.19844} {"mode": "train", "epoch": 72, "iter": 20, "lr": 0.00472, "memory": 6697, "data_time": 0.24143, "loss_prob": 0.97624, "loss_db": 0.16567, "loss_thr": 0.43851, "loss": 1.58042, "time": 1.26011} {"mode": "val", "epoch": 72, "iter": 80, "lr": 0.00472, "hmean-iou:recall": 0.7807, "hmean-iou:precision": 0.83962, "hmean-iou:hmean": 0.80909} {"mode": "train", "epoch": 73, "iter": 5, "lr": 0.00468, "memory": 6697, "data_time": 1.7954, "loss_prob": 0.85423, "loss_db": 0.15098, "loss_thr": 0.45753, "loss": 1.46274, "time": 2.90421} {"mode": "train", "epoch": 73, "iter": 10, "lr": 0.00468, "memory": 6697, "data_time": 0.29776, "loss_prob": 0.85731, "loss_db": 0.14573, "loss_thr": 0.39932, "loss": 1.40237, "time": 1.23691} {"mode": "train", "epoch": 73, "iter": 15, "lr": 0.00468, "memory": 6697, "data_time": 0.27035, "loss_prob": 0.85157, "loss_db": 0.14417, "loss_thr": 0.39404, "loss": 1.38979, "time": 1.33959} {"mode": "train", "epoch": 73, "iter": 20, "lr": 0.00468, "memory": 6697, "data_time": 0.24757, "loss_prob": 0.97287, "loss_db": 0.17615, "loss_thr": 0.4258, "loss": 1.57482, "time": 1.25033} {"mode": "val", "epoch": 73, "iter": 80, "lr": 0.00468, "hmean-iou:recall": 0.79825, "hmean-iou:precision": 0.80531, "hmean-iou:hmean": 0.80176} {"mode": "train", "epoch": 74, "iter": 5, "lr": 0.00465, "memory": 6697, "data_time": 1.77191, "loss_prob": 0.79204, "loss_db": 0.13724, "loss_thr": 0.40282, "loss": 1.33211, "time": 2.81048} {"mode": "train", "epoch": 74, "iter": 10, "lr": 0.00465, "memory": 6697, "data_time": 0.35505, "loss_prob": 0.90555, "loss_db": 0.16267, "loss_thr": 0.43164, "loss": 1.49986, "time": 1.32076} {"mode": "train", "epoch": 74, "iter": 15, "lr": 0.00465, "memory": 6697, "data_time": 0.28285, "loss_prob": 0.84884, "loss_db": 0.14985, "loss_thr": 0.40945, "loss": 1.40814, "time": 1.26473} {"mode": "train", "epoch": 74, "iter": 20, "lr": 0.00465, "memory": 6697, "data_time": 0.22423, "loss_prob": 0.83922, "loss_db": 0.13588, "loss_thr": 0.37352, "loss": 1.34863, "time": 1.17557} {"mode": "val", "epoch": 74, "iter": 80, "lr": 0.00465, "hmean-iou:recall": 0.80702, "hmean-iou:precision": 0.81416, "hmean-iou:hmean": 0.81057} {"mode": "train", "epoch": 75, "iter": 5, "lr": 0.00462, "memory": 6697, "data_time": 1.89862, "loss_prob": 0.85342, "loss_db": 0.15324, "loss_thr": 0.38652, "loss": 1.39318, "time": 3.14093} {"mode": "train", "epoch": 75, "iter": 10, "lr": 0.00462, "memory": 6697, "data_time": 0.40921, "loss_prob": 0.91439, "loss_db": 0.1647, "loss_thr": 0.41639, "loss": 1.49547, "time": 1.22508} {"mode": "train", "epoch": 75, "iter": 15, "lr": 0.00462, "memory": 6697, "data_time": 0.24701, "loss_prob": 0.86097, "loss_db": 0.16405, "loss_thr": 0.4155, "loss": 1.44053, "time": 1.17545} {"mode": "train", "epoch": 75, "iter": 20, "lr": 0.00462, "memory": 6697, "data_time": 0.21653, "loss_prob": 0.83336, "loss_db": 0.14021, "loss_thr": 0.41576, "loss": 1.38932, "time": 1.21775} {"mode": "val", "epoch": 75, "iter": 80, "lr": 0.00462, "hmean-iou:recall": 0.77193, "hmean-iou:precision": 0.85437, "hmean-iou:hmean": 0.81106} {"mode": "train", "epoch": 76, "iter": 5, "lr": 0.00459, "memory": 6697, "data_time": 1.8262, "loss_prob": 0.96722, "loss_db": 0.16866, "loss_thr": 0.42837, "loss": 1.56425, "time": 3.11078} {"mode": "train", "epoch": 76, "iter": 10, "lr": 0.00459, "memory": 6697, "data_time": 0.34228, "loss_prob": 0.83945, "loss_db": 0.13983, "loss_thr": 0.41511, "loss": 1.39439, "time": 1.24785} {"mode": "train", "epoch": 76, "iter": 15, "lr": 0.00459, "memory": 6697, "data_time": 0.24269, "loss_prob": 0.77103, "loss_db": 0.13846, "loss_thr": 0.39751, "loss": 1.30701, "time": 1.16602} {"mode": "train", "epoch": 76, "iter": 20, "lr": 0.00459, "memory": 6697, "data_time": 0.23284, "loss_prob": 0.82548, "loss_db": 0.14137, "loss_thr": 0.4183, "loss": 1.38514, "time": 1.17493} {"mode": "val", "epoch": 76, "iter": 80, "lr": 0.00459, "hmean-iou:recall": 0.80702, "hmean-iou:precision": 0.83636, "hmean-iou:hmean": 0.82143} {"mode": "train", "epoch": 77, "iter": 5, "lr": 0.00455, "memory": 6697, "data_time": 1.89037, "loss_prob": 0.86756, "loss_db": 0.14867, "loss_thr": 0.43502, "loss": 1.45125, "time": 3.01924} {"mode": "train", "epoch": 77, "iter": 10, "lr": 0.00455, "memory": 6697, "data_time": 0.28837, "loss_prob": 0.80347, "loss_db": 0.13739, "loss_thr": 0.42769, "loss": 1.36856, "time": 1.19132} {"mode": "train", "epoch": 77, "iter": 15, "lr": 0.00455, "memory": 6697, "data_time": 0.25167, "loss_prob": 0.77439, "loss_db": 0.13453, "loss_thr": 0.38776, "loss": 1.29668, "time": 1.23971} {"mode": "train", "epoch": 77, "iter": 20, "lr": 0.00455, "memory": 6697, "data_time": 0.23134, "loss_prob": 0.9148, "loss_db": 0.16201, "loss_thr": 0.43418, "loss": 1.51099, "time": 1.17831} {"mode": "val", "epoch": 77, "iter": 80, "lr": 0.00455, "hmean-iou:recall": 0.80702, "hmean-iou:precision": 0.85981, "hmean-iou:hmean": 0.83258} {"mode": "train", "epoch": 78, "iter": 5, "lr": 0.00452, "memory": 6697, "data_time": 1.92259, "loss_prob": 0.81544, "loss_db": 0.14349, "loss_thr": 0.43224, "loss": 1.39118, "time": 3.1056} {"mode": "train", "epoch": 78, "iter": 10, "lr": 0.00452, "memory": 6697, "data_time": 0.36711, "loss_prob": 0.79501, "loss_db": 0.13893, "loss_thr": 0.40264, "loss": 1.33658, "time": 1.271} {"mode": "train", "epoch": 78, "iter": 15, "lr": 0.00452, "memory": 6697, "data_time": 0.24121, "loss_prob": 0.80556, "loss_db": 0.14648, "loss_thr": 0.40289, "loss": 1.35494, "time": 1.27677} {"mode": "train", "epoch": 78, "iter": 20, "lr": 0.00452, "memory": 6697, "data_time": 0.2458, "loss_prob": 0.83969, "loss_db": 0.14757, "loss_thr": 0.41819, "loss": 1.40545, "time": 1.21414} {"mode": "val", "epoch": 78, "iter": 80, "lr": 0.00452, "hmean-iou:recall": 0.80702, "hmean-iou:precision": 0.80702, "hmean-iou:hmean": 0.80702} {"mode": "train", "epoch": 79, "iter": 5, "lr": 0.00449, "memory": 6697, "data_time": 2.18414, "loss_prob": 0.77072, "loss_db": 0.13662, "loss_thr": 0.40215, "loss": 1.3095, "time": 3.49511} {"mode": "train", "epoch": 79, "iter": 10, "lr": 0.00449, "memory": 6697, "data_time": 0.55483, "loss_prob": 0.80347, "loss_db": 0.13312, "loss_thr": 0.42606, "loss": 1.36265, "time": 1.4829} {"mode": "train", "epoch": 79, "iter": 15, "lr": 0.00449, "memory": 6697, "data_time": 0.14666, "loss_prob": 0.85638, "loss_db": 0.14485, "loss_thr": 0.42296, "loss": 1.42419, "time": 1.19862} {"mode": "train", "epoch": 79, "iter": 20, "lr": 0.00449, "memory": 6697, "data_time": 0.22317, "loss_prob": 0.83489, "loss_db": 0.13855, "loss_thr": 0.40241, "loss": 1.37585, "time": 1.24663} {"mode": "val", "epoch": 79, "iter": 80, "lr": 0.00449, "hmean-iou:recall": 0.72807, "hmean-iou:precision": 0.93258, "hmean-iou:hmean": 0.81773} {"mode": "train", "epoch": 80, "iter": 5, "lr": 0.00445, "memory": 6697, "data_time": 1.90082, "loss_prob": 0.81433, "loss_db": 0.14039, "loss_thr": 0.41021, "loss": 1.36494, "time": 3.28465} {"mode": "train", "epoch": 80, "iter": 10, "lr": 0.00445, "memory": 6697, "data_time": 0.26379, "loss_prob": 0.77511, "loss_db": 0.13023, "loss_thr": 0.4288, "loss": 1.33414, "time": 1.25493} {"mode": "train", "epoch": 80, "iter": 15, "lr": 0.00445, "memory": 6697, "data_time": 0.29805, "loss_prob": 1.04423, "loss_db": 0.1674, "loss_thr": 0.44752, "loss": 1.65915, "time": 1.29785} {"mode": "train", "epoch": 80, "iter": 20, "lr": 0.00445, "memory": 6697, "data_time": 0.2231, "loss_prob": 0.79777, "loss_db": 0.13595, "loss_thr": 0.41598, "loss": 1.3497, "time": 1.23199} {"mode": "val", "epoch": 80, "iter": 80, "lr": 0.00445, "hmean-iou:recall": 0.81579, "hmean-iou:precision": 0.80172, "hmean-iou:hmean": 0.8087} {"mode": "train", "epoch": 81, "iter": 5, "lr": 0.00442, "memory": 6697, "data_time": 1.99578, "loss_prob": 0.98923, "loss_db": 0.16665, "loss_thr": 0.42496, "loss": 1.58084, "time": 3.14785} {"mode": "train", "epoch": 81, "iter": 10, "lr": 0.00442, "memory": 6697, "data_time": 0.22379, "loss_prob": 0.95486, "loss_db": 0.16162, "loss_thr": 0.43368, "loss": 1.55015, "time": 1.23538} {"mode": "train", "epoch": 81, "iter": 15, "lr": 0.00442, "memory": 6697, "data_time": 0.37464, "loss_prob": 0.79042, "loss_db": 0.14499, "loss_thr": 0.40032, "loss": 1.33574, "time": 1.41556} {"mode": "train", "epoch": 81, "iter": 20, "lr": 0.00442, "memory": 6697, "data_time": 0.22999, "loss_prob": 0.80191, "loss_db": 0.13728, "loss_thr": 0.39115, "loss": 1.33035, "time": 1.23056} {"mode": "val", "epoch": 81, "iter": 80, "lr": 0.00442, "hmean-iou:recall": 0.76316, "hmean-iou:precision": 0.85294, "hmean-iou:hmean": 0.80556} {"mode": "train", "epoch": 82, "iter": 5, "lr": 0.00439, "memory": 6697, "data_time": 1.88433, "loss_prob": 0.8251, "loss_db": 0.14148, "loss_thr": 0.41003, "loss": 1.37661, "time": 3.11013} {"mode": "train", "epoch": 82, "iter": 10, "lr": 0.00439, "memory": 6697, "data_time": 0.52942, "loss_prob": 0.76283, "loss_db": 0.12862, "loss_thr": 0.42403, "loss": 1.31549, "time": 1.52728} {"mode": "train", "epoch": 82, "iter": 15, "lr": 0.00439, "memory": 6697, "data_time": 0.24408, "loss_prob": 0.77767, "loss_db": 0.13427, "loss_thr": 0.3993, "loss": 1.31124, "time": 1.20119} {"mode": "train", "epoch": 82, "iter": 20, "lr": 0.00439, "memory": 6697, "data_time": 0.23964, "loss_prob": 0.89766, "loss_db": 0.16041, "loss_thr": 0.41673, "loss": 1.4748, "time": 1.22605} {"mode": "val", "epoch": 82, "iter": 80, "lr": 0.00439, "hmean-iou:recall": 0.81579, "hmean-iou:precision": 0.80172, "hmean-iou:hmean": 0.8087}

[gaotongxiao]

Empirically, ~80 epochs might not be sufficient. You can keep training it for another few hundred epochs. Check out the log here

[Tomhouxin]

但是学习率已经下降到很低了

[Tomhouxin]

And detect result like this

[gaotongxiao]

The result looks reasonable to me. If you are striking for higher performance, consider trying out different data augmentation techniques, adding more data (e.g. using CTW1500), and different learning strategies. We cannot give concrete suggestions as the actual situation varies case by case.

[Tomhouxin]

Ok, Thank you！！！