WeijingShi
commented
2 years ago Hi @rajahaseeb147,
For any dataset, we need to two basic APIs: read_points(sample_id) and read_boxes(sample_id).
read_points() should return points and their attributes such as reflections.
read_boxes() should return the object class and its bounding boxes.
The points and the bounding box should be in the same coordinate frame. Otherwise, we need the calibration information to make them so.
In this Point-GNN version, we read point cloud data here:
https://github.com/WeijingShi/Point-GNN/blob/48f3d79d5b101d3a4b8439ba74c92fcad4f7cab0/train.py#L79
Besides some coordinate transformations, get_cam_points_in_image_with_rgb simply returns [N, 3] point coordinates and [N, 4] point attributes (intensity, r, g, b).
We read boxes here:
https://github.com/WeijingShi/Point-GNN/blob/48f3d79d5b101d3a4b8439ba74c92fcad4f7cab0/train.py#L81
It returns the object class and KITTI object boxes. Note, many autonomous driving datasets and detection frameworks use a 7-DoF box format (center_x, center_y, center_z, length, width, height, heading) assuming we only care about the vehicle's yaw angle. In your medical dataset, this type of box or any box, in general, may not be a good representation. Maybe segmentation is more suitable.
You can implement your own read_points(simple_id) and read_boxes(simple_id) and replace ours in the code. Or just make your dataset in KITTI's format. I think most of the frameworks support the KITTI format.
Hope it helps. Weijing
pytholic
Hi. I have been doing research regarding point cloud object detection. However, all the architectures currently support only benchmark datasets and there is no proper guidance on how to prepare our own datasets. For example, I have some point clouds of patients and I want to detect some regions for medical purposes. However, I cannot find any guidance here or in other models as well to work with custom datasets. Will appreciate any guidance on this. Thanks!