이 프로젝트는 딥러닝 기반 3D 객체 검출 결과물을 바탕으로, 차량의 Visible Faces 을 추정하여 3D Bounding Box의 위치를 정밀하게 보정하는 알고리즘을 구현한 것입니다.
- Base Paper: 3D Bounding Box Estimation Using Deep Learning and Geometry (arXiv:1612.00496)
- Base Repository: skhadem/3D-BoundingBox
This project focuses on refining the 3D Bounding Box positions generated by deep learning models. By building upon the initial orientation, translation, and dimensions estimated from the baseline model, our algorithm predicts the visible faces of surrounding vehicles to perform geometric refinement.
The system leverages the geometric relationship between the visible feature points and the vehicle's center, minimizing localization errors and providing more robust 3D object poses in complex environments.
보정 알고리즘의 핵심은 특징점으로부터 차량 중심까지의 오프셋을 계산하여 기하학적으로 정렬하는 것입니다.
- Orange Points: 오프셋 계산의 기준이 되는 특징점
- Red Lines: 특징점에서 차량 중심까지의 오프셋 벡터
- Green Boxes: 최종 보정된 3D Bounding Box (Proposed)
- White Boxes: Original deep learning results (Base Repository)
- Green Boxes: Refined results after our geometry-based correction
| 1. 시뮬레이터 차량 배치 (Simulator) | 2. 이미지 상의 3D BBox (Image) |
|---|---|
 |
 |
| 3. 3D BBox 탑뷰 (Top View) | 4. 3D BBox 프론트뷰 (Front View) |
|---|---|
 |
 |
If the bridge (e.g., CARLA-ROS bridge) does not broadcast TF data, please uncomment the robot_state_publisher node in your launch file to enable manual TF broadcasting.