How kuanchihhuang/ptt works
Project type and purpose
• This is a research‐oriented, open‐source deep-learning pipeline for temporal 3D object detection on LiDAR data. It extends the OpenPCDet framework with a Point-Trajectory Transformer (PTT) model targeting Waymo and other autonomous-driving datasets.
File structure
• Top-level directories:
• pcdet/ ← core library (config, datasets, models, ops, utils)
• tools/ ← entry‐point scripts (train.py, test.py, demo.py), configs, eval & visual utils
• docker/ ← containerization
• docs/ ← tutorials, changelog
• pcdet/datasets: loaders & augmentors for Waymo, KITTI, NuScenes, custom, etc.
• pcdet/models: backbones (2D/3D), VFE modules, dense heads, detector templates, ROI heads (including PTT head).
• pcdet/ops: custom CUDA/C++ kernels (IoU NMS, PointNet2, ROI‐aware pooling).
• tools/cfgs: YML files describing model hyperparameters and dataset paths.
• tools/train.py, test.py: orchestrate data loading, model instantiation, distributed training/inference.
README for additional insights
• Confirms PTT builds on OpenPCDet; requires pretraining CenterPoint RPN; then trains PTT using ROI proposals from earlier model.
• Workflow:
a) Preprocess raw Waymo data with OpenPCDet
b) Run CenterPoint (4 frames) → outputs ROI pickle files
c) Train PTT (32 frames) using those pickles
d) Evaluate/test with test.py
Final advice
• Every stage: data, model, ops, training scripts, environment.
• Do not over-abstract; each major folder is its own block.