train

[GXJll]

Dear author, I would like to consult how to make your own data set, and then use your own data set to train on your model.Thanks

[InhwanBae]

Hi @GXJll,

Thanks for your interest in our work!

1. Make your own dataset

Creating a user dataset for training is quite straightforward. First, make a new folder inside the dataset directory. Within it, add three subfolders: train, val, and test. Each of these should contain one or more .txt files describing pedestrian trajectories.

Each .txt file should store trajectory coordinate sequences in tab-seperated format, where each row corresponds to a single observation:

# Video frame \t Pedestrian ID \t x-coordinate \t y-coordinate
# Trajectory of pedestrian 1
780	1	8.46	3.59
790	1	9.57	3.79
800	1	10.67	3.99
810	1	11.73	4.32
820	1	12.81	4.61
# Trajectory of pedestrian 2
800	2	13.64	5.8
810	2	12.09	5.75
820	2	11.37	5.8
830	2	10.31	5.97
840	2	9.57	6.24
# (...)
1020	2	-1.52	6.05
# Trajectory of pedestrian 3
830	3	12.49	6.6
840	3	11.94	6.77
850	3	11.03	6.84
860	3	10.21	6.81
870	3	9.36	6.85
# (...)
1020	3	-0.72	6.66
# (...)

Here, video frames are downsampled to 10fps from a 25fps video. You can use any frame rate for your dataset, but it should be consistent for all scenes within the dataset. Coordinates are in meters, but any metrics like pixels are okay. For reference, please check biwi_eth.txt.

The folder structure should look like this:

./dataset
└─ user_dataset
   ├─ train
   │  ├─ scene_train.txt
   │  └─ scene_train_2.txt (optional)
   ├─ val
   │  └─ scene_val.txt
   │  └─ scene_val_2.txt (optional)
   └─ test
      ├─ scene_test.txt
      └─ scene_test_2.txt (optional)

2. Training the NPSN Model

Training NPSN involves two stages. The total training time is approximately 3 hours. First, train the backbone model (e.g., PECNet, STGCNN, or SGCN). For PECNet, run the following command:

cd ./baselines/pecnet/
python3 train_pecnet_ethucy.py --dataset user_dataset --gpu_num 0

Note

You can also try other backbone models by using their official training code (Social-STGCNN, SGCN).
They are fully compatible with your user_dataset since the dataloader is the same.

Once the backbone is trained, you can train the NPSN module with the following command:

./train_npsn.sh -b pecnet -d "user_dataset" -i "0"

[GXJll]

Thank you very much for answering my question. I still don't quite understand the method you mentioned. The picture shows a frame of my video dataset. Do you have any way to make it into a dataset and train it? Is there a video tutorial on YouTube? Could you provide it? Thank you very much and look forward to hearing from you.

[gxj980611-cyber]

Hello author, after a long period of exploration, I am now able to train PCENET backbone, but can other backbone networks also be trained? If so, I would like to ask for advice on how to train them

[InhwanBae]

Hi @GXJll,

Sorry for the late reply. You can't use the video as is in the 'camera' window. Instead, you can use the top-down view trajectory coordinate data to train our NPSN model, indicated by the arrow in the background image. I think there is trajectory coordinate data that is used to generate a synthetic 'camera' video. If you can access the trajectory coordinate data, it's surprisingly easy to convert it to the CSV format I've shown above.

[InhwanBae]

Hi @gxj980611-cyber

Thanks for your interest in our paper! If you want to use a different backbone like Social-STGCNN and SGCN with your own data, you'll first need to train it using their original repositories. Once you have your best pre-trained model, just drop the file into our pretrained folder. Our NPSN code will then automatically load it.

The training process in those other repos is very similar to ours, so you can use the same data to train the backbone.

[GXJll]

感谢作者回复，但是我在训练过程中，遇到了ADE和FDE一直为50的情况，不知道是哪里除了问题，您是否有何建议？下面是我自制的数据集格式，另外我加载训练好的模型，在预测时，在第一二帧数的时候预测的较为准确，后面的误差竟能到达200m不知是和原因，您是否有能够优化的建议。
1316 1 0.65 7.71
1317 1 0.29 7.74
1318 1 -0.03 7.75
1319 1 -0.39 7.74
1320 1 -0.76 7.72
1321 1 -1.14 7.68
1322 1 -1.50 7.63
1323 1 -1.90 7.54
1324 1 -2.31 7.43
1325 1 -2.72 7.30
1326 1 -3.13 7.14
1327 1 -3.54 6.95
1 2 0.98 8.37
2 2 0.95 8.37
3 2 0.91 8.38
4 2 0.90 8.40
5 2 0.87 8.40
6 2 0.80 8.40
7 2 0.70 8.40
8 2 0.57 8.40
9 2 0.42 8.41
10 2 0.26 8.42
11 2 0.10 8.43
12 2 -0.04 8.43
13 2 -0.22 8.43
14 2 -0.36 8.41
15 2 -0.48 8.41
16 2 -0.57 8.41
17 2 -0.66 8.42
18 2 -0.77 8.42
19 2 -0.92 8.41
20 2 -1.04 8.39

[InhwanBae]

Hi @GXJll,

I think your data looks almost perfect. Does the issue where ADE and FDE stay fixed at 50 occur during the training of the baseline model, or during the training of NPSN?

NPSN/baselines/pecnet/train_pecnet_ethucy.py

Lines 176 to 177 in 9d2fca3

	best_test_loss = 50 # start saving after this threshold
	best_endpoint_loss = 50

I think these lines are related to your issues. In the case of PECNet, the initial error is set to 50, and I used it exactly as it was in the original implementation. PECNet is very prone to overfitting, but this can be avoided with strong data augmentation. Please add the following augmentation at line 37 of train_pecnet_ethucy.py.

obs_traj = obs_traj + torch.randn_like(obs_traj) * 0.1     # adjust 0.1 to the optimal value
pred_traj = pred_traj + torch.randn_like(pred_traj) * 0.1  # adjust 0.1 to the optimal value

[InhwanBae]

Thank you very much, your suggestion is very effective and has been successfully resolved.

I'm glad to hear that! Feel free to reach out anytime if you need more assistance.

…

On Mon, Oct 13, 2025 at 11:32 AM, GXJll ***@***.***> wrote: *GXJll* left a comment (InhwanBae/NPSN#6) <#6 (comment)> Thank you very much, your suggestion is very effective and has been successfully resolved. — Reply to this email directly, view it on GitHub <#6 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ALATQKVJLC6HIRN5BFHRQF33XMFK7AVCNFSM6AAAAACF5J6UTKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTGOJVGY3TCMJZHE> . You are receiving this because you commented.Message ID: ***@***.***>