3D Object Detection Algorithm Based on Graph Neural Network and Dynamic Sampling

Table 1. Experimental environment
View table
Table 1. Experimental environment
Experimental environment Configuration
Operating system Ubuntu 22.04
CPU Intel Core i5-13490F
GPU GeForce RTX 4060ti （16 GB）
Deep learning framework PyTorch 1.11.0
Programming environment Python 3.8.0

Table 2. Comparison of AP_3D and inference speed of different algorithms on KITTI validation set

View table

Table 2. Comparison of AP_3D and inference speed of different algorithms on KITTI validation set

Algorithm	Car AP_3D			Pedestrian AP_3D			Cyclist AP_3D			Inference speed /（frame/s）
Algorithm	Easy	Moderate	Hard	Easy	Moderate	Hard	Easy	Moderate	Hard	Inference speed /（frame/s）
SECOND^［12］	88.52	79.09	76.21	49.75	44.46	39.85	81.28	65.13	60.64	38.46
PointPillars^［13］	88.59	77.01	73.95	48.81	43.28	39.28	82.12	62.40	58.32	40.98
IASSD^［10］	89.97	80.08	78.97	58.25	53.96	48.96	91.55	71.79	67.30	46.76
3DSSD^［9］	88.84	78.58	76.62	54.41	48.38	43.34	89.08	68.18	65.58	28.09
CoIn^［28］	87.69	78.49	76.81	53.87	50.67	46.85	85.12	67.46	63.35	35.97
Point-GNN^［17］	88.65	77.57	73.86	52.06	44.56	41.16	79.65	64.58	58.73	11.90
PV-RCNN^［29］	91.46	82.58	80.17	59.66	51.57	45.65	87.92	70.59	66.41	15.53
PointRCNN^［8］	89.75	80.56	78.06	60.72	53.38	47.87	89.78	70.43	66.89	17.54
Voxel R-CNN^［15］	90.28	81.91	80.30	61.53	54.15	49.01	89.67	71.50	67.54	20.16
DSVT-Voxel^［30］	92.16	81.25	80.02	60.23	54.36	49.81	89.46	72.69	67.83	17.42
Proposed	92.98	83.14	81.46	61.61	55.58	51.05	91.05	73.15	68.76	26.39

Table 3. Comparison of VRAM usage experiments
View table
Table 3. Comparison of VRAM usage experiments
Algorithm VRAM usage /MB GPU utilization /% Time /s mAP_3D /%
PointPillars 6939 89 223 63.75
Proposed 13560 96 463 73.20

Table 4. Results of algorithm ablation experiments

View table

Table 4. Results of algorithm ablation experiments

Algorithm	GFE	DPS	MGP	mAP_3D /%			mAP_BEV/%
Algorithm	GFE	DPS	MGP	Car	Pedestrian	Cyclist	Car	Pedestrian	Cyclist
Baseline	×	×	×	79.85	43.79	67.61	89.59	49.93	70.67
	√	×	×	81.83	47.37	68.61	90.61	53.30	71.99
	×	√	√	85.26	55.80	76.16	91.13	59.91	77.23
	√	×	√	85.10	54.42	73.72	91.81	60.13	78.17
	√	√	√	85.86	56.08	77.65	91.83	60.29	78.82

Table 5. Ablation experiment results of each component of GFE
View table
Table 5. Ablation experiment results of each component of GFE
VFE SADR FDFS mAP_3D /%
Car Pedestrian Cyclist
× × × 79.86 42.33 66.05
√ × × 80.23 44.56 66.92
√ √ × 80.69 45.26 67.56
√ √ √ 81.85 46.06 68.26

Table 6. Impact of hyperparameter K value in GFE
View table
Table 6. Impact of hyperparameter K value in GFE
Algorithm K Car AP_3D /%
Easy Moderate Hard
Proposed 9 92.87 83.12 81.42
16 91.34 82.69 80.57
32 89.35 82.79 80.69

Table 7. Comparison of performance and time consumption between DPS and traditional sampling methods
View table
Table 7. Comparison of performance and time consumption between DPS and traditional sampling methods
Method Car AP_3D /% Time consumption /ms
Overall 0‒20 m 20‒40 m 40 m‒inf
FPS 82.33 93.34 82.46 35.63 183
DPS 82.96 94.56 84.53 38.26 172

Table 8. Impact of parameter t on detection performance and inference speed
View table
Table 8. Impact of parameter t on detection performance and inference speed
t Pedestrian mAP_3D /% Cyclist mAP_3D /% Inference speed /（frame/s）
1.0 53.14 74.23 27.08
1.2 54.36 75.54 26.69
1.4 56.08 77.65 26.39
1.6 55.51 76.82 25.89
1.8 54.06 75.07 25.74

Table 9. Adaptive module ablation experiment
View table
Table 9. Adaptive module ablation experiment
Adaptive module Car AP_3D /% Car AP_BEV /%
× 85.16 91.23
√ 85.82 91.86

Tools

Get Citation

Copy Citation Text

Kai Zhong, Ying Chen, Chengzhi Yan, Han Gao. 3D Object Detection Algorithm Based on Graph Neural Network and Dynamic Sampling[J]. Laser & Optoelectronics Progress, 2025, 62(14): 1415003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Nov. 27, 2024

Accepted: Jan. 20, 2025

Published Online: Jul. 16, 2025

The Author Email: Ying Chen (chy@sit.edu.cn)

DOI:10.3788/LOP242327

CSTR:32186.14.LOP242327

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Experimental environment

Table 1. Experimental environment

Table 2. Comparison of AP3D and inference speed of different algorithms on KITTI validation set

Table 2. Comparison of AP3D and inference speed of different algorithms on KITTI validation set

Table 3. Comparison of VRAM usage experiments

Table 3. Comparison of VRAM usage experiments

Table 4. Results of algorithm ablation experiments

Table 4. Results of algorithm ablation experiments

Table 5. Ablation experiment results of each component of GFE

Table 5. Ablation experiment results of each component of GFE

Table 6. Impact of hyperparameter K value in GFE

Table 6. Impact of hyperparameter K value in GFE

Table 7. Comparison of performance and time consumption between DPS and traditional sampling methods

Table 7. Comparison of performance and time consumption between DPS and traditional sampling methods

Table 8. Impact of parameter t on detection performance and inference speed

Table 8. Impact of parameter t on detection performance and inference speed

Table 9. Adaptive module ablation experiment

Table 9. Adaptive module ablation experiment

Table 2. Comparison of AP_3D and inference speed of different algorithms on KITTI validation set

Table 2. Comparison of AP_3D and inference speed of different algorithms on KITTI validation set