핵심 개념
Using disk covering for real-time instance segmentation achieves state-of-the-art results with efficient resource utilization.
초록
I. Abstract
Increasing accuracy in instance segmentation often sacrifices speed.
Coarser representations reduce parameters for real-time masks.
II. Introduction
Intelligent vehicles rely on computer vision for environment interpretation.
Urban settings require real-time detection of road users.
Deep learning models are resource-intensive due to large parameter sizes.
III. Method
Based on CenterNet, predicts object centers and radii for disk covering.
Gaussian projection used for dense covering with N disks per pixel.
Training without custom ground-truth using binary masks.
IV. Experiments
Tested on Cityscapes, IDD, and KITTI datasets for road user segmentation.
Achieved state-of-the-art results with real-time inference on RTX 3090 GPU.
V. Discussion
Ablation studies on disk numbers, loss functions, and post-processing.
Limitations include difficulty in capturing straight lines and sharp angles.
통계
우리의 방법은 Cityscapes 테스트 세트에서 1 이미지당 0.040초의 추론 시간을 달성했습니다.
IDD 데이터 세트에서 AP는 20.30이고 AP50%는 49.90입니다.
KITTI 데이터 세트에서 AP는 11.75이고 AP50%는 37.24입니다.
인용구
"Our method achieve state-of-the art results on the IDD and KITTI dataset with an inference time of 0.040 s on a single RTX 3090 GPU."
"The best performances are met on the categories that are most present in the dataset, and that have the biggest surface in average."