Project Introduction: Det3D
Det3D is an advanced framework tailored for 3D object detection, primarily built using PyTorch. It provides comprehensive implementations of various 3D object detection algorithms and offers robust performance on major benchmark datasets.
Overview of Det3D
Det3D stands out as the first toolbox dedicated to 3D object detection. It simplifies the implementation process by offering ready-to-use versions of widely-known algorithms like PointPillars, SECOND, and PIXOR. Additionally, it supports state-of-the-art methods that are benchmarked against leading datasets such as KITTI (ViP) and nuScenes (CBGS).
Key Features:
- Multi-Dataset Compatibility: Det3D is compatible with multiple datasets, including KITTI, nuScenes, and Lyft.
- Extensive Model Zoo: It includes both point-based and voxel-based models.
- Top-tier Performance: The models maintain high performance on popular evaluation metrics.
- Advanced Training Techniques: Det3D supports distributed data parallel (DDP) training and synchronized batch normalization (SyncBN).
Getting Started with Det3D
The framework offers detailed guides for installation and a quick start through dedicated documentation, such as the INSTALLATION.md for setup guidance and GETTING_STARTED.md for a step-by-step initiation process.
Model Zoo Insights
Det3D hosts a variety of models fine-tuned for specific datasets:
nuScenes:
- Examples of models trained on nuScenes include CBGS, which delivers a mean Average Precision (mAP) of 49.9%, and PointPillar, achieving a mAP of 41.8%.
KITTI:
- On the KITTI dataset, the Second model yields impressive results with a bounding box AP of 90.54%.
- PointPillars on KITTI provides a similar level of performance with a bounding box AP of 90.63%.
Lyft and Waymo:
- Det3D also offers models configured for the Lyft and Waymo datasets, expanding its applicability across popular datasets.
Functionalities and Development
Det3D encompasses a wide range of functionalities to enhance 3D object detection:
Models Included:
- VoxelNet
- SECOND
- PointPillars
Features Encompass:
- Multi-task learning capabilities.
- Distributed Training and Validation for efficient resource use.
- Customizable anchor dimensions for flexibility.
- Supports TensorboardX for detailed experiment visualization.
Future Plans
The roadmap includes releasing additional models like PointRCNN and PIXOR, as well as expanding support for the Waymo Dataset. Furthermore, the team is seeking contributions to extend the list of supported models and datasets.
Contribution and Licensing
Det3D encourages contributions and is released under the Apache license. It is a derivative work based on the CBGS approach, which is well-regarded in the academic community.
Acknowledgements
Det3D builds upon various established projects and libraries such as mmdetection, mmcv, and maskrcnn_benchmark, benefiting from their robust foundations and widespread support.
For those interested in advanced 3D object detection, Det3D presents a valuable resource with substantial capabilities and community support.