Hello!
Thank you for sharing this amazing job.

I have a question about the code, and I am wondering that I use this code in the right way.
In 'sample.py' code for calculating ADE and FDE, does this code scale x,y position to real value?
This is because, in the preprocess step for generating .ckpl file, the preprocess function scales (x,y) position down to the range of (-1, 1).
Which means, if I want to get real predicted results, do I need to scale the position up?

In addition, I am wondering that numerical analysis of FDE and ADE is calculated by the position (x,y) which is not scaled up.

I trained the model, but when I am plotting the results, it looks like this

This is not nearly as good as the images they show in the paper you reference though. Do you have a pretrained model with decent results?

I have trained the model and when I run the sample.py I have plotted the output but I need to Visualize on the image can you help me how to visualize on the image

Thank You
Daniel

Thanks for putting this out there! Could you clarify the license and add a license file? Is it MIT?

Hi,
I have some doubts about the pre-processing part.
Why does it want to multiply by two and then reduce one ( *2 - 1 ) at "frame_preprocess"?
What is the relationship between this and normalization?

data[:, 3] = (
(data[:, 3] - min_position_x) / (max_position_x - min_position_x)
) * 2 - 1
data[:, 4] = (
(data[:, 4] - min_position_y) / (max_position_y - min_position_y)
) * 2 - 1

You are comparing your results with the results from the paper and report that your results are better. However, you are using the normalized values to calculate this error as far as I can see. Since the error depends on the scale, you assume that the paper did the same preprocessing of the data as you did (division by factor 2000 as far as I can see that from the code). How do you know that this is the same as the authors did?

When I run the train.py, this error occurs,would you help me to check this question?3Q
RuntimeError: size mismatch, m1: [2 x 3], m2: [64 x 64] at C:/w/1/s/tmp_conda_3.7_044431/conda/conda-bld/pytorch_1556686009173/work/aten/src\THC/generic/THCTensorMathBlas.cu:268

I am looking at the code and cannot figure out what the numbers in your labeled data file indicate. An explanation would be good.

Additionally, in the following picture, is the circled number an identifier for a vehicle/pedestrian/bicycle? If so, how is it numbered?

We are looking to retrain the model with different category identifiers above 3 (1 through 5). Is that feasible?

how I can get a graph from a photograph , like this paper ,how transform each frame of the picture into graph ?

hello，I have a problem，and in the trajectory prediction process, the data set is replaced by prediction-test, and a series of errors are always reported. In addition, it is known that predict 4 frames after 6 frames, why there are more than 6 frames of data in the prediction dataset. Thank you for you to give guidance and help

估计您是使用apollo的轨迹数据集，您所使用的数据集在https://github.com/ApolloScapeAuto/dataset-api/tree/master/trajectory_prediction有所介绍，主要是对Data Structure的部分信息不理解（相信作者对该数据集应该理解透彻了，希望能给予点解答）：
Each line in a file contains frame_id, object_id, object_type, position_x, position_y, position_z, object_length, object_width, object_height, heading. 这句话中heading是什么？还有position_x, position_y, position_z的坐标系的原点在哪以及x轴方向,y轴方向,z轴方向是哪？这个三维坐标如何转换到图片上的二维坐标？

After read the code of input, you pre-process the data (x, y) to [-1, 1]. So the error is calculated with normalized result and ground truth. So how to get absolute error?
Like Trajectory result of : http://apolloscape.auto/leader_board.html

some of the training loss are negative, how does this happen?