search

WeijingShi / Point-GNN

Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud, CVPR 2020.
MIT License
540 stars 112 forks source link

Field of view #25

Closed Juuustin closed 4 years ago

Juuustin commented 4 years ago

Hi, I test the algorithm on my dataset and find that the field of view is only about 120 degrees in front of the ego car. I am wondering if this algorithm could detect 360 degrees of view.

WeijingShi commented 4 years ago

Hi @Juuustin,

KITTI dataset only provides ground truth for the objects in front of the front camera. That's why the code removes the points outside the front image when loading the point cloud run.py Line 210. A quick hack would be replacing get_cam_points_in_image_with_rgb with get_cam_points to process all points. But since the checkpoints are not trained with objects outside the view, the results would likely be messy. If you have your own dataset, it's better to retrain the model in the complete view.

Juuustin commented 4 years ago

Hi @Juuustin,

KITTI dataset only provides ground truth for the objects in front of the front camera. That's why the code removes the points outside the front image when loading the point cloud run.py Line 210. A quick hack would be replacing get_cam_points_in_image_with_rgb with get_cam_points to process all points. But since the checkpoints are not trained with objects outside the view, the results would likely be messy. If you have your own dataset, it's better to retrain the model in the complete view.

Make sense, and thank you very much!