search

WeijingShi / Point-GNN

Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud, CVPR 2020.
MIT License
523 stars 114 forks source link

作者你好，我有个问题想请教下 #58

Closed EvaneSnow closed 3 years ago

EvaneSnow commented 3 years ago
  1. 你放在checkpoints中的权重是使用KITTI数据集训练的吗？
  2. 我还有必要使用KITTI数据集继续进行训练吗？
  3. 我想对单张图片进行预测，请问如何操作？ 希望得到您的回复，谢谢。
EvaneSnow commented 3 years ago

@WeijingShi 希望得到你的回复，谢谢

WeijingShi commented 3 years ago

Hi @EvaneSnow, Thanks for your interest in our work.

  1. Yes, the checkpoints are all trained on KITTI.
  2. Sure, you can explore better methods to re-train, fine-tuning or modify the model using KITTI data or other datasets.
  3. We use a point cloud as input instead of an image. The easiest method to run on your own data is to organize it in KITTI format and run the inference code: python3 run.py checkpoints/car_auto_T3_trainval/ --test --dataset_root_dir DATASET_ROOT_DIR --output_dir DIR_TO_SAVE_RESULTS.

Hope it helps.

EvaneSnow commented 3 years ago

@WeijingShi 谢谢你的回答，我看你的用户名应该也是中国人，我就不使用英语和你交流了，抱歉哈。我是刚接触点云预测网络，偶然看到你发表在CVPR上的论文，很有兴趣。我想问哈，如果我使用这条命令进行训练： python3 train.py configs/car_auto_T3_train_train_config configs/car_auto_T3_train_config 那么，我的训练类别是只有车这一类吗？能不能将车，人一起进行训练？ 我在你的car_auto_T3_train_train_config文件中发现，num_classes：4，这是说明有4类车吗？ 希望得到您的回复，谢谢。

WeijingShi commented 3 years ago

@EvaneSnow, Sure, no problem. Currently, we use separate checkpoints for cars and pedestrians, but you can try to combine them into one. The num_classes=4 contains the background, side-view car, front/back-view car, and other vehicles.

EvaneSnow commented 3 years ago

@WeijingShi 谢谢你的回答，如果我想在前向推理过程中，看到如你论文中的三维框，二维框， image

是不是需要设置python3 run.py checkpoints/car_auto_T3_train/ --dataset_root_dir DATASET_ROOT_DIR --output_dir DIR_TO_SAVE_RESULTS --level 1 将--level的值设置为1.

WeijingShi commented 3 years ago

Yes, that's correct. You can also set level to 2 to view the results one by one.

EvaneSnow commented 3 years ago

@WeijingShi 非常感谢你的回答，谢谢你，祝你学术事业双丰收。

EvaneSnow commented 3 years ago

@WeijingShi 我最后还有一个小问题，我看你论文中，前向推理，可以同时检测车，人 image 如果我也想在同一张点云上同时检测车，人，如何操作呢？ 因为我看项目中的权重，车和人的权重参数是分开保存的，那进行权重加载时，如何进行操作？

WeijingShi commented 3 years ago

Hi @EvaneSnow, The simplest way is to run two models independently on the same inputs and then put their results together. Of course, putting all classes into one network and re-train would be a better method. Best wishes to you as well!

EvaneSnow commented 3 years ago

真是太感谢你啦，我目前已经没有问题啦