search

HimangiM / Just-Go-with-the-Flow-Self-Supervised-Scene-Flow-Estimation

Self-supervised method for scene-flow estimation of LiDAR point clouds. Method is trained and tested on the nuScenes and KITTI datasets in TensorFlow. (CVPR 2020)
BSD 3-Clause "New" or "Revised" License
60 stars 5 forks source link

no file about model_concat_upsa_1nn_cycle_nuscenes.py ? #11

Open Haiyan-Chris-Wang opened 3 years ago

Haiyan-Chris-Wang commented 3 years ago

@HimangiM Hi, in your command_train_cycle_nuscenes.sh file, there seems an model named model_concat_upsa_1nn_cycle_nuscenes. However, we couldn't find it in your repo. Would you mind provide it later?

HimangiM commented 3 years ago

Hi,

Thank you for pointing it out. You can use the model named model_concat_upsa_cycle under the src folder.

Haiyan-Chris-Wang commented 3 years ago

@HimangiM Hi, thanks for answering the question. Just a few small questions and it would be appreciated a lot if you could help to answer. 1). In the paper, you mention that you are using the Flownet3d pre-trained model on the FlyingThing3D dataset. I am just wondering is it supervised pretraining? During this pretraining process, have you applied the flip strategy? 2). I am trying to rewrite your work with pytorch module and will release it later to contribute the community. But I found the knn_l2_loss could converge quickly but the cycle_l2_loss cannot converge at all. Do you have any idea of this problem?