The KITTI Vision Benchmark Suite

Method

Enhancing Monocular 3D Detection through Data Augmentation Strategies [DA3D]

Submitted on 23 Jun. 2023 04:26 by
贾译凇 (哈尔滨工业大学)

Running time:		0.03 s
Environment:		1 core @ 2.5 Ghz (Python)

Method Description:

We migrate the 2D object detection data augmentation method to 3D object detection without losing coordinate cues and size cues, and experimentally prove that our migration method is necessary. The sampling bias problem of data augmentation is explored in this process, and an unbiased sampling strategy and several new augmentation methods for 3D object detection are proposed, and the effectiveness of the proposed methods is experimentally verified.

Parameters:

lr=1e-4

Latex Bibtex:

Detailed Results

Object detection and orientation estimation results. Results for object detection are given in terms of average precision (AP) and results for joint object detection and orientation estimation are provided in terms of average orientation similarity (AOS).

Benchmark	Easy	Moderate	Hard
Car (Detection)	0.00 %	0.00 %	0.00 %
Car (Orientation)	0.00 %	0.00 %	0.00 %
Car (3D Detection)	27.76 %	20.47 %	17.89 %
Car (Bird's Eye View)	36.83 %	26.92 %	23.41 %
Pedestrian (Detection)	0.00 %	0.00 %	0.00 %
Pedestrian (Orientation)	0.00 %	0.00 %	0.00 %
Pedestrian (3D Detection)	4.62 %	2.95 %	2.58 %
Pedestrian (Bird's Eye View)	4.93 %	3.27 %	2.74 %
Cyclist (Detection)	0.00 %	0.00 %	0.00 %
Cyclist (Orientation)	0.00 %	0.00 %	0.00 %
Cyclist (3D Detection)	3.37 %	1.86 %	1.48 %
Cyclist (Bird's Eye View)	3.46 %	1.89 %	1.51 %