Method

vast-net [vast-net]
[Anonymous Submission]

Submitted on 1 Feb. 2018 14:33 by
[Anonymous Submission]

Running time:0.3 s
Environment:GPU @ 3.5 Ghz (Python)

Method Description:
Extracting features using the vast-net
Parameters:
score_threshold: 0.0
iou_threshold: 0.6
max_detections_per_class: 100
max_total_detections: 100
nms_score_threshold: 0.0
nms_iou_threshold: 0.7
max_proposals: 300
localization_loss_weight: 2.0
objectness_loss_weight: 1.0
initial_crop_size: 17
maxpool_kernel_size: 1
maxpool_stride: 1

RMSProp with decay of 0:9 and  = 1:0. We used a
learning rate of 0:045, decayed every two epochs
using
an exponential rate of 0:94
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) 87.56 % 87.35 % 78.32 %
Car (Orientation) 51.85 % 49.43 % 44.99 %
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




eXTReMe Tracker