Title Page
Contents
ABSTRACT 10
국문초록 12
CHAPTER 1. INTRODUCTION 14
CHAPTER 2. RELATED WORK 19
2.1. 3D object detection in autonomous driving 20
2.2. 3D object shape completion 22
2.3. Transformer-based in 3D object detection 23
CHAPTER 3. METHODOLOGY 25
3.1. Overall architecture 25
3.2. Rotation backbone 26
3.3. 3D object shape recovering 29
3.4. 3D detector based on attention mechanism. 30
3.4.1. Attention fusion mechanism. 30
3.4.2. Head prediction 33
3.5. Model loss function 34
3.5.1. 3D object classification loss 34
3.5.2. 3D object orientation loss 34
3.5.3. 3D object bounding box loss 35
3.5.4. Total loss 35
CHAPTER 4. EXPERIMENTS 36
4.1. KITTI 3D object detection dataset 36
4.2. Detail of implementation 40
4.2.1. Training details of the transformer shape 40
4.2.2. Training details of the 3D detector 40
CHAPTER 5. RESULTS 41
5.1. Evaluation metrics 41
5.2. Comparative analysis of transformer shape efficacy 41
5.3. Comparison with TED model 42
5.4. Comparison of cutting-edge methods 43
CHAPTER 6. CONCLUSION 46
REFERENCES 47
[Table 2-1] Survey 3D object detection method 19
[Table 4-1] Description of annotations for 3D object detection 39
[Table 5-1] Comparing the effects of transformer shape 42
[Table 5-2] Performace comparision with TED model 43
[Table 5-3] Compare with cutting-edge method 44
[Figure3-1] The overal architeture of our 3D Transformer Detection. 26
[Figure3-2] Perspective attention mechanism. 32
[Figure4-1] Visualization of KITTI dataset LiDAR data for 3d object detection. 38
[Figure5-1] The visualization of evaluation results on the KITTI Dataset. The model has demonstrated strong performance in detecting vehicle objects with... 45