如何仅使用1个标签选择正确的对象检测模型的图像？
用户酶：我正在尝试提取屏幕截图的某些部分，该部分是从游戏（带有TF对象检测模型）中取出的，并在此部分中提取文本（游戏中使用的字体自定义模型）。
我已经训练了基于SSD Mobilenet V2的自定义模型，并且对象检测可以很好地运行，但是有时界框关闭了。我搜索了选择合适的图像和适量训练自定义模型的金额，但找不到正确的提示。
我尝试提取以下内容（被红色包围）：
环境可以改变：
- 游戏的分辨率可以不同（1920x1080，whqd等），
- 盒子中的文本并不总是
与我接受了120张自行制作图像（1920x1080）训练（训练10％的90％） ％用于测试）（所有这些图像中的屏幕截图），正如我提到的，结果还不错。有时，检测区域关闭（切割盒子的内容或包含周围盒子周围的很多区域）。
也许有人可以帮助我/回答以下问题：
- 更大的培训数据集可以提高准确性吗？
- 创建培训数据时，我还应该考虑不同的决议吗？
- 只能在没有游戏屏幕截图的情况下只喂盒子，这是有意义的吗？还是我应该混合整个游戏的屏幕截图和仅框屏幕截图？
预先感谢您！ :)
UseCase: I'm trying to extract certain parts of a screenshot which is taken from a game (with a tf object detection model) and extract the text within this part (custom model for the font used in the game).
I have trained a custom model based on SSD Mobilenet V2 and the object detection works quite okish, but sometimes the bounding box is off. I googled about selecting the right images and the right amount for training the custom model, but I couldn't find a good hint in the right direction.
I try to extract the following (surrounded by red):
The environmen can change:
- Resolution of the game can be different (1920x1080, WHQD etc.)
- Text in the box is not always the same
I have trained with 120 self made images (1920x1080) (90% for training 10% for test) (all of these images where a screenshot of the game) and as I mentioned the results are okish. Sometimes the detected area is off (cutting the content of the box or including a lot area of the box surroundings).
Maybe someone can help me/answering the following questions:
- Could a bigger training dataset increase the accuracy?
- Should I also take different resolutions into account when creating the training data?
- Would it make sense to feed only the boxes without the rest of the game screenshot into the training? Or should I mix screenshots of the whole game and only box screenshots?
Thank you in advance ! :)
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论