Method

Stereo Neural Vernier Caliper [st] [SNVC]
https://github.com/Nicholasli1995/SNVC

Submitted on 26 Mar. 2022 08:54 by
Zechun Liu (HKUST)

Running time:1 s
Environment:GPU @ 1.0 Ghz (Python)

Method Description:
SNVC introduces multi-resolution 3D space modeling
into voxel-based stereo 3D object detection. A new
instance-level analysis model is proposed, which
samples candidate 3D locations and uses predicted
object part coordinates to estimate a pose update.
Parameters:
Mainscale network voxel size 0.2m
Latex Bibtex:
@inproceedings{li2022stereo,
title={Stereo Neural Vernier Caliper},
author={Li, Shichao and Liu, Zechun and Shen,
Zhiqiang and Cheng, Kwang-Ting},
booktitle={Proceedings of the AAAI Conference on
Artificial Intelligence},
volume={36},
year={2022}
}

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) 96.33 % 93.32 % 85.81 %
Car (Orientation) 96.27 % 93.09 % 85.51 %
Car (3D Detection) 78.54 % 61.34 % 54.23 %
Car (Bird's Eye View) 86.88 % 73.61 % 64.49 %
This table as LaTeX


2D object detection results.
This figure as: png eps pdf txt gnuplot



Orientation estimation results.
This figure as: png eps pdf txt gnuplot



3D object detection results.
This figure as: png eps pdf txt gnuplot



Bird's eye view results.
This figure as: png eps pdf txt gnuplot




eXTReMe Tracker