I want to finetune the trainable pointpillar with a custom dataset. I know the format of the annotation must be KITTI but does the object position refer to the center of the 3D bounding box or the bottom center ?
When preparing your label files (.txt files) for a custom dataset to be used with TAO PointPillars, you follow the standard KITTI dataset annotation format. In this format, the 3D position location for each object refers to the coordinates of the 3D bounding box’s bottom center in the camera coordinate system .
NVIDIA’s TAO PointPillars model performs its internal calculations, feature extraction, and object detection using the LiDAR coordinate system . The origin of this system is the center of the LiDAR sensor .
Before model training begins, the TAO Toolkit’s data loading and preprocessing pipeline (e.g., when generating .pkl files with the dataset_convert command) automatically handles the coordinate transformation. It reads the labels you provide in the KITTI format (based on the camera coordinate system and bottom-center reference) and uses the vehicle’s calibration files to transform this annotation data from the camera coordinate system into the LiDAR coordinate system that the model uses internally.
There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.