Method

Monocular 3D RetinaNet [RetinaMono]
[Anonymous Submission]

Submitted on 10 Dec. 2020 22:56 by
[Anonymous Submission]

Running time:0.02 s
Environment:1 core @ 2.5 Ghz (Python)

Method Description:
Our method extends highly accurate one-stage 2D object detector RetinaNet by adopting 3D anchor boxes that defines the location, shape, and orientation of potential objects in 3D real space. For the stereo setup, we also introduce an efficient approach that aggregates features of the original (left) image with those extracted from the additional (right) image at multiple depths sparsely and uniformly sampled to better localize the object-of-interest in 3D real space.
Parameters:
\alpha = 0.5, \beta = 0.5, \gamma = 2.0, 3-layer head
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) 75.18 % 69.01 % 58.98 %
Car (Orientation) 31.39 % 28.68 % 24.70 %
Car (3D Detection) 16.68 % 11.61 % 9.57 %
Car (Bird's Eye View) 24.52 % 16.85 % 14.02 %
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