oliverrensu / rtnet Goto Github PK

1.您好，我在测试test.py过程中发现DAVIS_dataset = RTTestDataset(dataset, 2, 384, int(384 * 1.75)),其中的dataset不知道如何组织；
2.查看RTTestDataset参数含义发现dataset为以下形式pathes使用：
for path in pathes:
file = sorted(os.listdir(os.path.join(path, "img")))
for i in file:
self.img_list.append(os.path.join(path, "img", i))
self.label_list.append(os.path.join(path, "label", i[:-3] + "png"))
self.fwflow_list.append(os.path.join(path, "flow_img", "fw_" + i[:-3] + "png"))
self.bwflow_list.append(os.path.join(path, "flow_img", "bw_" + i[:-3] + "png"))

其中的img，label等含义不是很清晰，希望能够得到您的帮助，非常感谢

您好，我发现您给出的分割结果是两种，有一种通过CRF进行了后处理，我使用您的预训练模型，可以实现未使用CRF的分割结果，但是我用CRF处理了这部分结果，效果不是特别好，我想请问一下这部分的实现细节是什么样的，十分感谢！

Could you please provide details of training the model of appearance (spatial-R34 or spatial RX-50) and motion stream (temporal-R34 or temporal RX-50)?
At first, I tested the pre-trained whole model on our custom dataset. The results are reasonable.

And then, I trained the Interactive model with the pre-trained spatial and temporal models on our custom dataset, the test results are still acceptable.

But when I train the whole model from scratch with our data, the performance is pretty poor.

Hello, I want to know why the flipped datasets are four times larger than the original (why not twice), and whether the original data and the flipped data are input together during the training process (Thanks for providing all the training datasets if it's convenient). Also, is it possible to improve the performance by inputting forward and backward OF together? Thanks again!

Thanks for your excellent work. However, when I try to train the model, the training process stops at this point.

Thanks for helping me out with this issue.

Thanks for your good work. The H and W of the input img of this function SelfScaling are 1 respectively? Otherwise, mul cannot be executed.

I would like to know how did you get the youtube-objects dataset? ?