Method

DUO-Net [DUO-Net]


Submitted on 5 May. 2025 04:55 by
Fazal ghaffar (Deakin University)

Running time:0.1 s
Environment:1 core @ 2.5 Ghz (Python)

Method Description:
In this work, we present DUO-Net, a unified multi-task learning framework for joint 2D object
detection and depth estimation. The architecture employs a shared ResNet-based backbone with
attention modules and task-specific heads to simultaneously perform bounding box localization
and dense depth prediction. We adopt a two-stage training strategy involving task-specific
pretraining followed by joint optimization to enhance representation learning and stability.
Parameters:
alpha=0.0
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) 96.19 % 95.24 % 90.60 %
Pedestrian (Detection) 71.70 % 62.48 % 59.97 %
Cyclist (Detection) 88.22 % 75.79 % 69.08 %
This table as LaTeX

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



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



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




eXTReMe Tracker