Junhao Huang

1. The unit of all above parameters is the length of images. For example, 0.05 means 30 pixels in a 600x600 image and 25.6 pixels in a 512x512 image. And...

I guess you are using `master` branch of **MarkToolForParkingLotPoint**, which does not annotate the direction and the type of marking-point. Please use [`directional_point` branch of **MarkToolForParkingLotPoint**](https://github.com/Teoge/MarkToolForParkingLotPoint/tree/directional_point) to label data for...

Good idea. Here is a link of our paper. I'll add it to readme soon. https://github.com/Teoge/DMPR-PS/blob/master/DMPR-PS.pdf

请把完整的 `Traceback` 发出来

Here `marks` stands for a directional parking-slot marking-point, and its four values represent its x, y coordinates, direction, and marking-point type respectively. You can use [https://github.com/Teoge/MarkToolForParkingLotPoint/blob/directional_point/mat2json.m](https://github.com/Teoge/MarkToolForParkingLotPoint/blob/directional_point/mat2json.m) to generate json file...

Multi-GPU training with pytorch is pretty simple. All you need to do is wrap the model with `nn.DataParallel`. Here are some links that might help you: [https://pytorch.org/tutorials/beginner/blitz/data_parallel_tutorial.html](https://pytorch.org/tutorials/beginner/blitz/data_parallel_tutorial.html) [https://pytorch.org/tutorials/beginner/former_torchies/parallelism_tutorial.html](https://pytorch.org/tutorials/beginner/former_torchies/parallelism_tutorial.html) Although I...

1. Resize 到 512 是因为神经网络的输入是 512x512。 2. PS2.0 数据集图片大小为 600x600，坐标减去 300.5 是将坐标原点置于图片中心，方便后续做 rotation augmentation。

> 好的，谢谢。 还想再请问一下为什么训练的时候在512x512大小的图片上面，而在推理的时候在原图大小上也可以得到准确的位置？ 因为神经网络输出的预测是 normalize 到 [0, 1] 的值，然后与图片大小相乘才得到具体的像素坐标

The labels are generated by my annotation tool [MarkToolForParkingLotPoint](https://github.com/Teoge/MarkToolForParkingLotPoint/tree/directional_point). No, they are not generated automatically, they require manual labelling. But if you are talking about other things than labelling, like...

Please refer to #16 .